- Email: [email protected]
Preoperative implant selection for unilateral breast reconstruction using 3D imaging with the Microsoft Kinect sensor
Accepted Manuscript Preoperative implant selection for unilateral breast reconstruction using 3D imaging with the Microsoft Kinect sensor Stefanie T.L. Pöhlmann, Elaine Harkness, Christopher J. Taylor, Ashu Gandhi, Susan M. Astley PII:
S1748-6815(17)30158-4
DOI:
10.1016/j.bjps.2017.04.005
Reference:
PRAS 5291
To appear in:
Journal of Plastic, Reconstructive & Aesthetic Surgery
Received Date: 17 December 2016 Revised Date:
31 March 2017
Accepted Date: 14 April 2017
Please cite this article as: Pöhlmann STL, Harkness E, Taylor CJ, Gandhi A, Astley SM, Preoperative implant selection for unilateral breast reconstruction using 3D imaging with the Microsoft Kinect sensor, British Journal of Plastic Surgery (2017), doi: 10.1016/j.bjps.2017.04.005. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT Preoperative implant selection for unilateral breast reconstruction using 3D imaging with the Microsoft Kinect sensor
1
RI PT
Stefanie T L Pöhlmann1, Elaine Harkness1, Christopher J Taylor1, Ashu Gandhi2, Susan M Astley1
Division of Informatics, Imaging & Data Sciences, Faculty of Biology, Medicine and Health,
2
SC
University of Manchester, Stopford Building, Oxford Road, Manchester M13 9PT, UK
University of Manchester, Manchester Academic Health Sciences Centre, University Hospital of
Please send correspondence to Susan M Astley
M AN U
South Manchester, Southmoor Road, Wythenshawe, Manchester M23 9LT, UK
Division of Informatics, Imaging & Data Sciences
TE D
School of Health Sciences Faculty of Biology, Medicine and Health University of Manchester
Oxford Road
EP
Stopford Building
Manchester M13 9PT
AC C
Email: [email protected] Telephone: +44 161 275 5162 Fax: +44 161275 5145
Parts of this work have been presented as a student presentation at the student and annual symposium of the Manchester Breast Centre.
1
ACCEPTED MANUSCRIPT Abstract Aims: This study aimed to investigate whether breast volume measured preoperatively with a Kinect 3D sensor could be used to determine the most appropriate implant size for reconstruction. Methods: Ten patients underwent 3D imaging before and after unilateral implant-based
RI PT
reconstruction. Imaging used seven configurations, varying patient pose and Kinect location, which were compared regarding suitability for volume measurement. Four methods of defining the breast boundary for automated volume calculation were compared, and repeatability assessed over five
SC
repetitions.
Results: The most repeatable breast boundary annotation used an ellipse to track the inframammary
M AN U
fold and a plane describing the chest wall (coefficient of repeatability: 70 ml). The most reproducible imaging position comparing pre- and postoperative volume measurement of the healthy breast was achieved for the sitting patient with elevated arms and Kinect centrally positioned (coefficient of repeatability: 141 ml). Optimal implant volume was calculated by correcting used implant volume by
TE D
the observed postoperative asymmetry. It was possible to predict implant size using a linear model derived from preoperative volume measurement of the healthy breast (coefficient of determination R2=0.78, standard error of prediction 120 ml). Mastectomy specimen weight and experienced surgeons’ choice showed similar predictive ability (both: R2=0.74, standard error: 141/142 ml). A
EP
leave one-out validation showed that in 61% of cases 3D imaging could predict implant volume to
AC C
within 10%, however for 17% of cases it was >30%. Conclusion: This technology has the potential to aid reconstruction surgery planning and implant procurement in order to maximize symmetry after unilateral reconstruction.
Keywords: Breast reconstruction, 3D imaging, Kinect, Implant selection, Mastectomy
2
ACCEPTED MANUSCRIPT Introduction The most common form of immediate breast reconstruction in the UK uses a tissue expander/implant1. Surgeons choose implants based on patient anatomy as assessed by subjective linear measurements2,3, personal experience and availability of implants. However, linear
RI PT
measurements including height, width and projection are insufficient to describe the breast shape and size accurately, and small measurement discrepancies may lead to variation in volumetric implant size estimation4 and potentially unacceptable asymmetry. Registry data shows that asymmetry after
SC
reconstruction necessitating further surgery affects around one fifth of patients5,6. An accurate and reproducible pre-operative measurement of breast size and shape would provide an objective
M AN U
assessment and help avoid asymmetry. Although various methods have been developed, there is no universally accepted objective method for determining breast volume7–9 and choosing implant size for breast reconstruction.
Recently, 3D imaging has found application in planning breast surgery including augmentation procedures10–14, reduction15 and reconstruction surgery16–20. Most studies used commercial 3D
TE D
scanning equipment such as the Konica Minolta Vivid21, Axis Three11 or Di3D22 surface scanners. Accuracy is very high (errors <1 mm), scanning speed is fast (<3 s), but commercial equipment is
use23.
EP
very expensive (20,000–130,000 US Dollars) and usually requires calibration at setup or before each
AC C
The use of 3D scanning with the Microsoft Kinect, a small sensor based on infra-red (non-ionising) light emission and originally designed for computer gaming has been investigated as a convenient, inexpensive alternative24–26. The technique is based on transportable and cheap equipment (~100 US Dollars), which is ready to use without time and labour intense installation and calibration. Recently the feasibility of using the Kinect for breast imaging has been demonstrated assessing cosmetic outcomes after breast surgery24,25 although it has not been used for planning breast reconstruction. Moreover, there is no consensus which imaging position is the most appropriate for this task, either in terms of arm position26 or body inclination27,28.In addition, Xi et al identified the definition of the
3
ACCEPTED MANUSCRIPT breast boundary and the of the posterior wall as critical factors that could potentially limit the accuracy of breast volume measurement7.
RI PT
Aims of this study This study aimed to analyse preoperative breast volume in patients with breast cancer, and hence predict the implant size for reconstruction. Different definitions of the breast boundary were studied to maximise measurement repeatability. Additionally, the most useful imaging position was determined
Methods Patient selection and 3D image acquisition
M AN U
SC
optimising reproducibility and predictive power to estimate implant sizes correctly.
For this explorative study, ten female patients who underwent unilateral mastectomy with breast reconstruction between March 2015 and May 2016 were enrolled. Eligible were patients without
TE D
history of breast surgery and whose reconstruction was performed as immediate implant-based reconstruction or two-stage procedure using a tissue expander. The Kinect II (Kinect for Xbox One, Microsoft) was used to acquire 3D images of the patients’ breasts before surgery (pre-operative) and
EP
at least two weeks after surgery (post-operative). A simple acquisition protocol was applied, omitting stationary equipment, as in 22,27, in addition to avoiding complicated patient positioning16,27 and time-
AC C
consuming fusion of multiple images25. The imaging protocol for both time points featured 7 different imaging positions, which combined different patient poses including sitting and lying with arms hanging and above the head with different Kinect positions (see Table 1). For imaging position 4, the Kinect was moved around the patient and a 3D model reconstructed continuously using open source software (KinectFusion)29. The treating surgeons were blinded to the image-based breast volume measurements and chose implants according to their current practice.
4
ACCEPTED MANUSCRIPT
Breast volume measurement All acquired 3D images were pre-processed by down-sampling using Meshlab (3D image points closer than 0.5 mm were merged) in order to guarantee fast and robust volume calculations. To
RI PT
separate the breast volume from the residual torso, breast boundaries were defined by manually selecting landmarks on the computer-generated 3D model surface, again using Meshlab. Four different boundary definitions were assessed for this feasibility study (see Figure 1):
SC
A. Three landmarks were annotated outlining the extreme points of the breast, indicated by the measurement points used by Qiao30. Inferiorly the lowest point of the inframammary fold
M AN U
was selected, medially, a point on the mid-sternal line at the level of maximum breast projection, and laterally, a point along the lateral projection of the inframammary fold at the same horizontal level as the medial landmark was chosen (the axillary fold was found to be unreliable and dependent on arm position). The internal boundary of the breast along the
TE D
chest wall was modelled by a flat plane intersecting all three extreme points of the breast. B. Landmarks were positioned tracking the inframammary fold. Landmarks were placed with at least one, every 30° with respect to the centre of the breast. If a clear infra-mammary fold was absent, a line of highest surface curvature was followed along the inferior aspect of the
EP
breast. The chest wall boundary was modelled by a flat plane, which was fitted to the
AC C
annotated points using least-squares fitting. C. Twelve points were annotated describing the “base of the breast” where the gland elevated from the chest wall as an ellipse. The ellipse was drawn, using a screen marker and desktop annotation program, first tracking the inframammary fold or a line of high curvature as surrogate using a stylus on a tablet PC. Then the less defined boundary segments were approximated by the elliptical shape constraint. Twelve points were subsequently placed every 30° along the ellipse. The chest wall was approximated using a flat plane fitted to the annotated points.
5
ACCEPTED MANUSCRIPT D. The same elliptical boundary annotation as C was used, but in contrast to A, B and C, the chest wall was not approximated by a flat plane. A curved surface described by thin-plate splines was used to model the chest wall19,31. The volume of the breast was calculated automatically based on the selected boundary definition
RI PT
from each 3D image using purpose-built software in Matlab32. Data evaluation
Patient demographics such as age, Body Mass Index (BMI) and bra size were documented and
SC
analysed, together with used implant sizes and mastectomy specimen weight. The relationship of excised tissue weight and implant chosen by the surgeon was explored as intraoperative specimen
M AN U
weight is sometimes seen as closely related to breast volume and can potentially serve as an approximate guide for reconstruction surgery33. Correlation was investigated by calculating the Spearman’s rank coefficient ρ. The coefficient of determination R2 34 was used to assess fit of a linear model based on weight to predict the chosen implants and the standard error comparing weight-based
TE D
implant selection and volume of the used implants, was calculated
The performance of 3D breast imaging for planning reconstructive surgery was evaluated, firstly investigating repeatability of the different boundary definitions (A, B, C, D) using the pre-surgery 3D
EP
images. The breast boundary was annotated five times for each patient using each of the four boundary definitions. Imaging position 5 (which covers the unaffected breast) was chosen for this
AC C
experiment as none of these images show mesh truncations, which can occur for the frontal Kinect position and potentially leaves the lateral aspect of very large breasts undescribed. In order to choose the most reliable boundary definition, Intra-Class Correlation (ICC)35 and the coefficient of repeatability36 were evaluated. The ICC relates the variance within repetitions ( ) with the variance
between subjects ( ): =
; repetitions for 95% of subjects are expected within one
. coefficient of repeatability: 2.77
The best imaging position was found, evaluating reproducibility with respect to how well patients adapted the same position in two sessions and the resulting influence on volume measurements. The
6
ACCEPTED MANUSCRIPT breast volume of the unaffected, healthy breast measured during both sessions was compared by calculating Spearman’s rank correlation and the coefficient of repeatability together with the 95% limits of agreement37.. 95% of the measurement discrepancies lie within the limits of agreement estimated as the mean discrepancy ± 1.96 standard deviations (of the discrepancies).
RI PT
Finally, power of preoperative 3D imaging to determine the optimal implant volume was assessed. The optimal implant volume ( ) was calculated by measuring present asymmetry after the reconstruction ( , −
! " #, ),
and correcting the used implant size (! ) by
SC
the offset: = ! + ( , −
! " #, )
M AN U
A linear model was used to describe the relationship between this optimal, corrected implant volume and the preoperative image-based volume measurement of the healthy breast. Predictive power was analysed by the coefficient of determination and compared with a model using specimen weight to predict optimal implant volume and a model taking into account both preoperatively measured breast
TE D
volume and specimen weight. Similarly, the relationship between the implant chosen by the experienced surgeon and the calculated implant size was assessed. The difference between predicted and optimal implant size was assessed in a leave one out fashion;
EP
i.e. the model was generated using all, but one, patients; implant size for the remaining patient was
AC C
predicted based on the model. This procedure was then repeated for all patients.
Results
Patient/ Intervention characteristics All patients (median age 51, range 34–61; median BMI 27, range 17–34; bra size range A–G cup) were treated with mastectomy for cancer (eight invasive carcinoma, two ductal cancer in situ). Seven patients in this study underwent skin-sparing mastectomy, three patients, skin and nipple sparing interventions. Eight patients were treated with immediate implant- based reconstruction, two patients
7
ACCEPTED MANUSCRIPT underwent a two-step procedure. All postoperative imaging was performed between 15 and 94 days after surgery, one patient was lost to follow-up. Implants used were silicone gel filled Mentor CPG™ 300 series implants, implant volumes used ranged from 155 to 620 ml (median 452 ml). Mastectomy specimens weighted between 165 and 988 g
RI PT
(median 396 g). Ease of use
The Kinect sensor was straight-forward to use in a clinical setting, even in small rooms. The
SC
acquisition protocol was simple and one imaging session with all seven 3D images took less than 30 minutes. Motion range in one patient was still limited after surgery and all poses with arms up were
M AN U
modified to be performed with arms elevated only within the pain-free motion range. The postoperative imaging, however, was, in all patients, used for result assessment only, not for implant size prediction.
Relationship between resection weight and used implant size
TE D
Weight of the excised specimen and chosen implant size were not significantly correlated (ρ=0.66, p=0.06). A coefficient of determination of R2=0.54 and a standard error of 200 ml comparing weight and implant size, illustrated the limited power of intraoperative weight measurement to predict
EP
implant selection (Figure 2).
AC C
Boundary annotation and definition of the chest wall For boundary annotation C, 95% of repeat volume measurements lay within 70 ml (= coefficient of repeatability). Variation among repeats was very small, compared to overall variation within the dataset (ICC = 0.996). Results were similar, but slightly worse for definition D, which used the same annotations as C, but approximated the chest wall as curved surface. Boundary definition A and B produced less stable results and were not used for following evaluations (coefficient of repeatability 167 ml for A,151 ml for B) (Table 2). Imaging Position
8
ACCEPTED MANUSCRIPT The most reproducible pose was achieved, when the patient sat upright with both arms above the head and hands resting on the head, the Kinect centrally positioned in front of the patient (95% of repeats within 141 ml and highest rank correlation comparing both imaging sessions ρ=0.93). The most variability was found for imaging position 2, where the patient was asked to sit upright and hold
RI PT
hands behind the back, which allowed some shoulder posture variation. Correlation between used implant volume and volume measured from the reconstructed breast was strongest for imaging position 6 and 7 (both ρ=0.78) (Table 3). Results for boundary annotation D were similar (coefficient
occurred for half of imaging positions (full results not shown).
M AN U
Implant choice
SC
of repeatability 110–307 ml) to method C, however, outliers, with maximum discrepancy >300 ml
When assessing the relationship between calculated optimal implant volume and the image-based volume of the healthy breast before the surgery, imaging position 5 produced the highest degree of correlation (ρ=0.88). Measurements taken from imaging position 4 showed the best fit of a linear model (R2=0.78) and most reliable prediction of implant size using the fitted model. Comparing
TE D
predicted implant sizes with the calculated ideal implant sizes resulted in a standard error of 120 ml. The weight and the choice of experienced surgeons resulted in similar, but lower coefficients of
(Figure 3).
EP
determination with the calculated optimal implant size (both: R2=0.74, standard error: 142/141 ml)
AC C
The patient with the largest breasts in our dataset exhibited the greatest discrepancy (37%) between selected implant size (620 ml) and optimal calculated implant size (982 ml for imaging position 4); a second patient showed >150 ml discrepancy, which equals an asymmetry of 100% in this case (Figure 4).
In order to access the quality of implant predictions using 3D imaging a leave-one-out validation was performed using images from image position 4: Data from 8 women was used to create a linear model, implant prediction based on this model was calculated for the 9th, which was left out for model creation. This was repeated for all datasets. In 61% of all configurations, discrepancy between optimal
9
ACCEPTED MANUSCRIPT implant volume and predicted implant volume was <10%; however, for 17% of configurations discrepancy was >30%.
RI PT
Discussion It was feasible to use the Kinect to measure preoperative breast volumes of cancer patients within less than 30 minutes. Using only one pose would reduce measurement time even further; we estimate less than 15 minutes are necessary including equipment setup.
SC
The most repeatable definition of the breast boundary was achieved, estimating the breast base as an ellipse (method C). This allowed indicating the breast boundary, where it is the most defined,
M AN U
including along the inframammary fold, and to extrapolate the boundary into segments where the breast is less clearly defined, such as superiorly, using the constraints of an ellipse. No clear benefits were found in modelling the internal boundary of the breast along the chest wall as a curved surface. This method produced several outliers, and may be less suited for clinical practice.
TE D
Potentially, body surface around the breast did not serve as a good indication of the shape of the chest wall, especially in patients with high BMI.
According to our measurements, it was most suitable to acquire 3D images of the patient’s breasts
EP
with implant prediction in mind, from 45° caudal angulation, either with the Kinect moving around the patient using KinectFusion (imaging position 4) or taking separate images for left and right breasts
AC C
from this position (imaging position 5/6). This is in agreement with the findings of Henseler and coworkers27, who suggested a tilted position of the patient instead of the imaging equipment. The patient should be seated with arms elevated resting on the head, in line with the finding that imaging position 3 was most reproducible. The use of a custom-build sensor holder for position 4 or 5/6 could potentially lead to similar reproducibility, eliminating uncertainties of using the Kinect in a hand-held fashion. Generally, reproducibility was likely to be underestimated in this study as some patients presented with visible weight change at the postoperative imaging session, which is not uncommon due to anxiety and/or commencing hormone therapy38.
10
ACCEPTED MANUSCRIPT Given the most suitable breast boundary definition and patient pose, we found preoperative imagebased volume measurement of the healthy breast indicated the optimal breast implant size with similar certainty to mastectomy specimen weight and to experienced surgeons. This technology could therefore aid planning for reconstruction surgery and implant procurement. Alternatively, the acquired
RI PT
3D models could enable the manufacture of customised implants, which is becoming possible with the advent of novel technologies such as 3D bio-printing 39–41.
SC
Conclusion
This study showed the feasibility of using fast, simple and inexpensive 3D imaging technology for
M AN U
predicting implant size before surgery, although there were significant technical challenges in determining breast volume by surface imaging.
A larger study is needed to further evaluate the potential of this technology to ensure breast cancer
Conflict of interest: None
EP
Ethical approval
TE D
patients get the optimum breast implants and best cosmetic outcome.
The study was approved by the NHS Research Ethics Committee Preston under the IRAS project
AC C
ID162918 and performed in accordance with the ethical standards of the 1964 Declaration of Helsinki and its subsequent amendments. The study was reported adhering to the STROBE guidelines. Acknowledgements
We would like to thank the nursing staff of the Nightingale Centre and the breast surgeons; Cliona Kirwan, James Harvey, John Murphy, Richard Johnson and Sumohan Chatterjee. We would also like to thank the charity Breast Cancer Now, which funded this study as part of a doctoral research studentship. The sponsors had no involvement in planning, conducting or publishing this study.
11
ACCEPTED MANUSCRIPT BIBILIOGAPHY
1.
Jeevan R, Cromwell D, Browne J, et al. National Mastectomy and Breast Reconstruction
2.
RI PT
Audit. NHS Inf. Cent. 2009. Tebbetts JB. A system for breast implant selection based on patient tissue characteristics and implant-soft tissue dynamics. Plast. Reconstr. Surg. 2002;109(4):1396-1409. doi:10.1097/00006534-200204010-00032.
Brown TP, Ringrose C, Hyland RE, Cole AA, Brotherston TM. A method of assessing female
SC
3.
doi:10.1054/bjps.1999.3110. 4.
M AN U
breast morphometry and its clinical application. Br. J. Plast. Surg. 1999;52(5):355-359.
Westreich M. Anthropomorphic breast measurement: protocol and results in 50 women with aesthetically perfect breasts and clinical application. Plast. Reconstr. Surg. 1997;100(2):46879. doi:10.1097/00006534-199708000-00032.
Hvilsom GB, Hölmich LR, Steding-Jessen M, et al. Delayed breast implant reconstruction: A
TE D
5.
10-year prospective study. J. Plast. Reconstr. Aesthetic Surg. 2011;64(11):1466-1474. doi:10.1016/j.bjps.2011.06.026.
Spear SL, Murphy DK, Slicton A, Walker PS. Inamed silicone breast implant core study
EP
6.
AC C
results at 6 years. Plast. Reconstr. Surg. 2007;120(7 Suppl 1):8S-16S; discussion 17S-18S. doi:10.1097/01.prs.0000286580.93214.df.
7.
Xi W, Perdanasari AT, Ong Y, et al. Objective breast volume, shape and surface area
assessment: A systematic review of breast measurement methods. Aesthetic Plast. Surg.
2014;38(6):1116-1130. doi:10.1007/s00266-014-0412-5. 8.
Kayar R, Civelek S, Cobanoglu M, Gungor O, Catal H, Emiroglu M. Five methods of breast volume measurement: a comparative study of measurements of specimen volume in 30 mastectomy cases. Breast cancer basic Clin. Res. 2011;5:43-52. doi:10.4137/BCBCR.S6128.
12
ACCEPTED MANUSCRIPT 9.
Choppin SB, Wheat JS, Gee M, Goyal A. The accuracy of breast volume measurement methods: A systematic review. The Breast 2016;28:121-129. doi:10.1016/j.breast.2016.05.010.
10.
Roostaeian J, Adams WP. Three-dimensional imaging for breast augmentation: Is this technology providing accurate simulations? Aesthetic Surg. J. 2014;34(6):857-875.
11.
RI PT
doi:10.1177/1090820X14538805. Mailey B, Freel A, Wong R, Pointer DT, Khoobehi K. Clinical accuracy and reproducibility of Portrait 3D surgical simulation platform in breast augmentation. Aesthetic Surg. J.
12.
SC
2013;33(1):84-92. doi:10.1177/1090820X12469807.
Donfrancesco A, Montemurro P, Hedén P. Three-dimensional simulated images in breast
M AN U
augmentation surgery. Plast. Reconstr. Surg. 2013;132(4):810-822. doi:10.1097/PRS.0b013e3182a014cb. 13.
Kovacs L, Eder M, Zimmermann A, et al. Three-dimensional evaluation of breast augmentation and the influence of anatomic and round implants on operative breast shape
14.
TE D
changes. Aesthetic Plast. Surg. 2012;36(4):879-887. doi:10.1007/s00266-012-9892-3. Gladilin E, Gabrielova B, Montemurro P, Hedén P. Customized planning of augmentation mammaplasty with silicon implants using three-dimensional optical body scans and
EP
biomechanical modeling of soft tissue outcome. Aesthetic Plast. Surg. 2011;35(4):494-501. doi:10.1007/s00266-010-9642-3. Eder M, Grabhorn A, Waldenfels F V., et al. Prediction of breast resection weight in reduction
AC C
15.
mammaplasty based on 3-dimensional surface imaging. Surg. Innov. 2013;20(4):356-364. doi:10.1177/1553350612460127.
16.
Szychta P, Raine C, Butterworth M, et al. Preoperative implant selection for two stage breast reconstruction with 3D imaging. Comput. Biol. Med. 2014;44:136-143. doi:10.1016/j.compbiomed.2013.09.013.
17.
Ahcan U, Bracun D, Zivec K, Pavlic R, Butala P. The use of 3D laser imaging and a new breast replica cast as a method to optimize autologous breast reconstruction after mastectomy.
13
ACCEPTED MANUSCRIPT The Breast 2012;21(2):183-189. doi:10.1016/j.breast.2011.09.009. 18.
Tomita K, Yano K, Hata Y, Nishibayashi A, Hosokawa K. DIEP flap breast reconstruction using 3-dimensional surface imaging and a printed mold. Plast. Reconstr. Surg. - Glob. Open 2015;3(3). doi:10.1097/GOX.0000000000000288. Tepper OM, Karp NS, Small K, et al. Three-dimensional imaging provides valuable clinical
RI PT
19.
data to aid in unilateral tissue expander-implant breast reconstruction. Breast J. 2008;14(6):543-50. doi:10.1111/j.1524-4741.2008.00645.x.
Tanabe YN, Honda T, Nakajima Y, Sakurai H, Nozaki M. Intraoperative application of three-
SC
20.
dimensional imaging for breast surgery. Scand. J. Plast. Reconstr. Surg. Hand Surg.
21.
M AN U
2005;39(6):349-352. doi:10.1080/02844310500296347.
Kovacs L, Eder M, Hollweck R, et al. Comparison between breast volume measurement using 3D surface imaging and classical techniques. The Breast 2007;16(2):137-45. doi:10.1016/j.breast.2006.08.001.
Catherwood T, McCaughan E, Greer E, Spence RAJ, McIntosh SA, Winder RJ. Validation of
TE D
22.
a passive stereophotogrammetry system for imaging of the breast: A geometric analysis. Med. Eng. Phys. 2011;33(8):900-905. doi:10.1016/j.medengphy.2011.02.005. Chae MP, Rozen WM, Spychal RT, Hunter-smith DJ. Breast volumetric analysis for aesthetic
EP
23.
planning in breast reconstruction: a literature review of techniques. Gland Surg.
24.
AC C
2015;5(6):212-226. doi:10.3978/j.issn.2227-684X.2015.10.03.
Oliveira HP, Silva MD, Magalhães A, Cardoso MJ, Cardoso JS. Is Kinect depth data accurate
for the aesthetic evaluation after breast cancer surgeries? In: Pattern Recognition and Image Analysis, LNCS7887. Springer Berlin Heidelberg; 2013:261-268.
25.
Lacher RM, Hipwell JH, Williams NR, Keshtgar MRS, Hawkes DJ, Stoyanov D. Low-cost surface reconstruction for aesthetic results assessment and prediction in breast cancer surgery. Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS 2015;2015-Novem:5871-5874. doi:10.1109/EMBC.2015.7319727.
14
ACCEPTED MANUSCRIPT 26.
Henseler H, Bonkat SK, Vogt PM, Rosenhahn B. The Kinect recording system for objective three- and four-dimensional breast assessment with image overlays. J. Plast. Reconstr. Aesthetic Surg. 2016;69(2):e27-e34. doi:10.1016/j.bjps.2015.10.021.
27.
Henseler H, Ju X, Ayoub A, Ray AK. The importance of the pose in three-dimensional
RI PT
imaging of the ptotic breast. J. Plast. Reconstr. Aesthetic Surg. 2013;66(11):1551-1556. doi:10.1016/j.bjps.2013.06.040. 28.
Reece GP, Merchant F, Andon J, et al. 3D surface imaging of the human female torso in
doi:10.1016/j.medengphy.2015.01.011.
Newcombe RA, Davison AJ, Izadi S, et al. KinectFusion: Real-time dense surface mapping
M AN U
29.
SC
upright to supine positions. Med. Eng. Phys. 2015;37(4):375-383.
and tracking. 10th IEEE Int. Symp. Mix. Augment. Real. 2011:127-136. doi:10.1109/ISMAR.2011.6092378. 30.
Qiao Q, Zhou G, Ling Y. Breast volume measurement in young Chinese women and clinical
31.
TE D
applications. Aesthetic Plast. Surg. 1997;21(5):362-368.
Kovacs L, Eder M, Hollweck R, et al. New aspects of breast volume measurement using 3dimensional surface imaging. Ann. Pastic Surg. 2006;57(6):602-10.
32.
EP
doi:10.1097/01.sap.0000235455.21775.6a.
Pöhlmann STL, Hewes J, Williamson AI, et al. Breast volume measurement using a games
AC C
console input device. In: Fujita H, Hara T, Muramatsu C, eds. IWDM 2014, LNCS 8539. Gifu: Springer International Publishing; 2014:666-673.
33.
Parmar C, West M, Pathak S, Nelson J, Martin L. Weight versus volume in breast surgery: an
observaParmar, C., West, M., Pathak, S., Nelson, J., & Martin, L. (2011). Weight versus volume in breast surgery: an observational study. JRSM Short Reports, 2(11), 87. http://doi.org/10.1258/shorts.2011.011070ti. JRSM Short Rep. 2011;2(11):87. doi:10.1258/shorts.2011.011070. 34.
Nagelkerke NJD. A note on a general definition of the coefficient of determination. Biometrika
15
ACCEPTED MANUSCRIPT 1991;78(3):691-692. doi:10.1093/biomet/78.3.691. 35.
Shrout PE, Fleiss JL. Intraclass correlations: Uses in assessing rater reliability. Psychol. Bull. 1979;86(2):420-428. doi:10.1037/0033-2909.86.2.420.
36.
Vaz S, Falkmer T, Passmore AE, Parsons R, Andreou P. The Case for Using the Repeatability
RI PT
Coefficient When Calculating Test-Retest Reliability. PLoS One 2013;8(9):1-7. doi:10.1371/journal.pone.0073990. 37.
Bland JM, Altman DG. Measuring agreement in method comparison studies. Stat. Methods
38.
SC
Med. Res. 1999;8(2):135-160. doi:10.1191/096228099673819272.
Vagenas D, DiSipio T, Battistutta D, et al. Weight and weight change following breast cancer:
M AN U
evidence from a prospective, population-based, breast cancer cohort study. BMC Cancer 2015;15(1):28. doi:10.1186/s12885-015-1026-2. 39.
Gerstle TL, Ibrahim AMS, Kim PS, Lee BT, Lin SJ. A Plastic Surgery Application in Evolution. Plast. Reconstr. Surg. 2014;133(2):446-451.
40.
TE D
doi:10.1097/01.prs.0000436844.92623.d3.
Tourassi GD, Vargas-Voracek R, Catarious DM, Floyd CE. Computer-assisted detection of mammographic masses: a template matching scheme based on mutual information. Med. Phys.
Rengier F, Mehndiratta A, Von Tengg-Kobligk H, et al. 3D printing based on imaging data:
AC C
41.
EP
2003;30:2123-2130. doi:10.1118/1.1589494.
Review of medical applications. Int. J. Comput. Assist. Radiol. Surg. 2010;5(4):335-341. doi:10.1007/s11548-010-0476-x.
16
ACCEPTED MANUSCRIPT
FIGURE LEGENDS Figure 1: Breast boundary definitions (A-D) applied to 3D image of a large breast (G-cup) with visible ptosis and pronounced inframammary fold and a small breast (A-cup) with absent ptosis, which made it
RI PT
necessary to estimate to inframammary fold, where curvature is highest
Figure 2: Relationship between resection weight and chosen implant volume (only nine points visible as data points overlap), the regression line is plotted as solid line, the line of equality is depicted as dashed
M AN U
SC
line
Figure 3: Relationship between image-based volume measurement, weight and implant choice by the surgeons with the calculated optimal implant volume (one patient was lost to follow-up, one dataset was incomplete lacking imaging position 4)
TE D
Figure 4: Patients with the largest discrepancy between used implant volume and calculated optimal implant volume, (a) largest volume discrepancy, (b) largest percent discrepancy; Kinect images pre- and post-surgery are shown; imaging position 03 allows judging asymmetry, for (b) a view from below is
AC C
EP
added, as asymmetry in the frontal view only mildly visible
17
ACCEPTED MANUSCRIPT
TABLES Table 1: The imaging protocol comprises seven imaging positions with different patient poses and Kinect locations 2
3
4
5
SC
Example
Patient
Sitting, arms
pose
arms back hanging
M AN U
Sitting, Sitting,
6
Sitting
Sitting
Sitting
arms above
arms
arms
head
above head
above head
arms up,
hands on
TE D
arms hanging
45° from
45° from
Moving in
the midline
the midline
On tripod,
On tripod,
an arc
and 45° of
and 45° of
Kinect
directly in
directly in
directly in
around the
cranial
cranial
location
front of
front of
front of
torso, 45°
angulation
angulation
subject
subject
of cranial
centred at
centred at
angulation
healthy
affected
breast
breast
EP
semi-
position,
On tripod,
AC C
Lying,
recumbent
head
subject
7
RI PT
1
45° of cranial angulation
18
ACCEPTED MANUSCRIPT
Table 2: Assessment of repeatability comparing sets of five repeat boundary annotations on each selected 3D image (most repeatable method shaded) Differences of repeats from patient mean [ml]
2
Max
correlation3
repeatability [ml]4
0.984
166.5
0.984
150.5
0.996
70.0
0.994
73.5
A
35.7
60.1
185.3
B
33.4
54.3
167.0
C
15.8
24.9
65.9
D
15.2
26.5
107.7
RI PT
Mean
SD
Absolute mean over all patients, for each patient the mean absolute difference was calculated, for the five individual
M AN U
1
Coefficient of
SC
1
Method
Intra-class Within subject
measurements with respect to the mean of those 5 repeats 2
) Calculated as square root of within-subject variance from a one-way ANOVA ('
3
Calculated as
4
Calculated as 2.77 ∙ '
()*
is the within-subject variance, ' is the between-groups variance from a one-way ANOVA , where '
AC C
EP
TE D
*
*
19
ACCEPTED MANUSCRIPT
Table 3: Assessment of reproducibility comparing volume of the healthy breast measured during the two imaging sessions for each imaging positions (best performance shaded) Spearman’s
Limits of Differences pre-post [ml]
Spearman’s
Coefficient
rank
of
rank
RI PT
agreement [ml]
correlation
correlation
Imaging Mean1
SD2
Max
lower
(post-
upper (pre-post)
position
repeatability 3
[ml]
implant)
108.6
269.1
-162.1
263.8
2
103.2
126.7
-266.8
-287.6
209.1
3
56.0
71.0
-152.4
-117.1
165.6
4
4.7
89.0
-122.4
-167.7
5/6
82.0
100.1
163.9
7
12.7
82.8
116.9
0.90
224.0
0.58
0.88
246.3
0.56
0.93
141.4
0.54
176.1
0.93
172.0
0.47
-195.3
197.3
0.91
183.5
0.78
-149.6
175.0
0.88
164.4
0.78
SC
80.5
M AN U
1
Absolute mean difference between pre and post-surgery imaging over all patients
2
Calculated from the signed difference (pre-post) over all patients
3
, where ' is the within-subject variance from a one-way ANOVA Calculated as 2.77 ∙ '
AC C
EP
TE D
1
20
ACCEPTED MANUSCRIPT
Table 4: Assessment of the relationship between optimal calculated implant volume and preoperative 3D imaging (best performance shaded) Spearman’s rank correlation (preoperative
Coefficient of determination (based on
position
volume measurement - optimal calculated
linear model predicting optimal implant
implant volume)
volume from preoperative volume
RI PT
Imaging
measurement)-
0.40
3
0.47
4
0.76
5/6
0.88
7
0.40
0.08
SC
2
0.35
0.42
M AN U
0.57
0.78
0.73 0.20
AC C
EP
TE D
1
21
AC C
EP
TE D
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT
1000 800 600 400
RI PT
Implant volume [cc]
1200
y = 0.341x + 252.47 R² = 0.5379
200 0
200
400
600
800
1000
1200
SC
0
AC C
EP
TE D
M AN U
Resection weight [g]
1200 1000 800 600
y = 1.1239x - 32.31 R² = 0.7773
400 200 0 0
200
400
600
800
1000
1200
SC
1000 800
y = 1.6415x - 282.7 R² = 0.7444
600 400 200 0 0
200
M AN U
1200
TE D
Calculated optimal implant volume [ml]
Implant volume derived form 3D imaging [ml]
400
600
800
1000
1200
Implants chosen by experienced surgeon [ml]
EP
1200 1000
AC C
Calculated optimal implant volume [ml]
RI PT
Calculated optimal implant volume [ml]
ACCEPTED MANUSCRIPT
800
y = 0.9567x + 125.69 R² = 0.7406
600 400 200 0
0
200
400
600
800
Specimen weight [g]
1000
1200
ACCEPTED MANUSCRIPT (a)
AC C
EP
TE D
M AN U
SC
RI PT
(b)
S1748-6815(17)30158-4
DOI:
10.1016/j.bjps.2017.04.005
Reference:
PRAS 5291
To appear in:
Journal of Plastic, Reconstructive & Aesthetic Surgery
Received Date: 17 December 2016 Revised Date:
31 March 2017
Accepted Date: 14 April 2017
Please cite this article as: Pöhlmann STL, Harkness E, Taylor CJ, Gandhi A, Astley SM, Preoperative implant selection for unilateral breast reconstruction using 3D imaging with the Microsoft Kinect sensor, British Journal of Plastic Surgery (2017), doi: 10.1016/j.bjps.2017.04.005. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT Preoperative implant selection for unilateral breast reconstruction using 3D imaging with the Microsoft Kinect sensor
1
RI PT
Stefanie T L Pöhlmann1, Elaine Harkness1, Christopher J Taylor1, Ashu Gandhi2, Susan M Astley1
Division of Informatics, Imaging & Data Sciences, Faculty of Biology, Medicine and Health,
2
SC
University of Manchester, Stopford Building, Oxford Road, Manchester M13 9PT, UK
University of Manchester, Manchester Academic Health Sciences Centre, University Hospital of
Please send correspondence to Susan M Astley
M AN U
South Manchester, Southmoor Road, Wythenshawe, Manchester M23 9LT, UK
Division of Informatics, Imaging & Data Sciences
TE D
School of Health Sciences Faculty of Biology, Medicine and Health University of Manchester
Oxford Road
EP
Stopford Building
Manchester M13 9PT
AC C
Email: [email protected] Telephone: +44 161 275 5162 Fax: +44 161275 5145
Parts of this work have been presented as a student presentation at the student and annual symposium of the Manchester Breast Centre.
1
ACCEPTED MANUSCRIPT Abstract Aims: This study aimed to investigate whether breast volume measured preoperatively with a Kinect 3D sensor could be used to determine the most appropriate implant size for reconstruction. Methods: Ten patients underwent 3D imaging before and after unilateral implant-based
RI PT
reconstruction. Imaging used seven configurations, varying patient pose and Kinect location, which were compared regarding suitability for volume measurement. Four methods of defining the breast boundary for automated volume calculation were compared, and repeatability assessed over five
SC
repetitions.
Results: The most repeatable breast boundary annotation used an ellipse to track the inframammary
M AN U
fold and a plane describing the chest wall (coefficient of repeatability: 70 ml). The most reproducible imaging position comparing pre- and postoperative volume measurement of the healthy breast was achieved for the sitting patient with elevated arms and Kinect centrally positioned (coefficient of repeatability: 141 ml). Optimal implant volume was calculated by correcting used implant volume by
TE D
the observed postoperative asymmetry. It was possible to predict implant size using a linear model derived from preoperative volume measurement of the healthy breast (coefficient of determination R2=0.78, standard error of prediction 120 ml). Mastectomy specimen weight and experienced surgeons’ choice showed similar predictive ability (both: R2=0.74, standard error: 141/142 ml). A
EP
leave one-out validation showed that in 61% of cases 3D imaging could predict implant volume to
AC C
within 10%, however for 17% of cases it was >30%. Conclusion: This technology has the potential to aid reconstruction surgery planning and implant procurement in order to maximize symmetry after unilateral reconstruction.
Keywords: Breast reconstruction, 3D imaging, Kinect, Implant selection, Mastectomy
2
ACCEPTED MANUSCRIPT Introduction The most common form of immediate breast reconstruction in the UK uses a tissue expander/implant1. Surgeons choose implants based on patient anatomy as assessed by subjective linear measurements2,3, personal experience and availability of implants. However, linear
RI PT
measurements including height, width and projection are insufficient to describe the breast shape and size accurately, and small measurement discrepancies may lead to variation in volumetric implant size estimation4 and potentially unacceptable asymmetry. Registry data shows that asymmetry after
SC
reconstruction necessitating further surgery affects around one fifth of patients5,6. An accurate and reproducible pre-operative measurement of breast size and shape would provide an objective
M AN U
assessment and help avoid asymmetry. Although various methods have been developed, there is no universally accepted objective method for determining breast volume7–9 and choosing implant size for breast reconstruction.
Recently, 3D imaging has found application in planning breast surgery including augmentation procedures10–14, reduction15 and reconstruction surgery16–20. Most studies used commercial 3D
TE D
scanning equipment such as the Konica Minolta Vivid21, Axis Three11 or Di3D22 surface scanners. Accuracy is very high (errors <1 mm), scanning speed is fast (<3 s), but commercial equipment is
use23.
EP
very expensive (20,000–130,000 US Dollars) and usually requires calibration at setup or before each
AC C
The use of 3D scanning with the Microsoft Kinect, a small sensor based on infra-red (non-ionising) light emission and originally designed for computer gaming has been investigated as a convenient, inexpensive alternative24–26. The technique is based on transportable and cheap equipment (~100 US Dollars), which is ready to use without time and labour intense installation and calibration. Recently the feasibility of using the Kinect for breast imaging has been demonstrated assessing cosmetic outcomes after breast surgery24,25 although it has not been used for planning breast reconstruction. Moreover, there is no consensus which imaging position is the most appropriate for this task, either in terms of arm position26 or body inclination27,28.In addition, Xi et al identified the definition of the
3
ACCEPTED MANUSCRIPT breast boundary and the of the posterior wall as critical factors that could potentially limit the accuracy of breast volume measurement7.
RI PT
Aims of this study This study aimed to analyse preoperative breast volume in patients with breast cancer, and hence predict the implant size for reconstruction. Different definitions of the breast boundary were studied to maximise measurement repeatability. Additionally, the most useful imaging position was determined
Methods Patient selection and 3D image acquisition
M AN U
SC
optimising reproducibility and predictive power to estimate implant sizes correctly.
For this explorative study, ten female patients who underwent unilateral mastectomy with breast reconstruction between March 2015 and May 2016 were enrolled. Eligible were patients without
TE D
history of breast surgery and whose reconstruction was performed as immediate implant-based reconstruction or two-stage procedure using a tissue expander. The Kinect II (Kinect for Xbox One, Microsoft) was used to acquire 3D images of the patients’ breasts before surgery (pre-operative) and
EP
at least two weeks after surgery (post-operative). A simple acquisition protocol was applied, omitting stationary equipment, as in 22,27, in addition to avoiding complicated patient positioning16,27 and time-
AC C
consuming fusion of multiple images25. The imaging protocol for both time points featured 7 different imaging positions, which combined different patient poses including sitting and lying with arms hanging and above the head with different Kinect positions (see Table 1). For imaging position 4, the Kinect was moved around the patient and a 3D model reconstructed continuously using open source software (KinectFusion)29. The treating surgeons were blinded to the image-based breast volume measurements and chose implants according to their current practice.
4
ACCEPTED MANUSCRIPT
Breast volume measurement All acquired 3D images were pre-processed by down-sampling using Meshlab (3D image points closer than 0.5 mm were merged) in order to guarantee fast and robust volume calculations. To
RI PT
separate the breast volume from the residual torso, breast boundaries were defined by manually selecting landmarks on the computer-generated 3D model surface, again using Meshlab. Four different boundary definitions were assessed for this feasibility study (see Figure 1):
SC
A. Three landmarks were annotated outlining the extreme points of the breast, indicated by the measurement points used by Qiao30. Inferiorly the lowest point of the inframammary fold
M AN U
was selected, medially, a point on the mid-sternal line at the level of maximum breast projection, and laterally, a point along the lateral projection of the inframammary fold at the same horizontal level as the medial landmark was chosen (the axillary fold was found to be unreliable and dependent on arm position). The internal boundary of the breast along the
TE D
chest wall was modelled by a flat plane intersecting all three extreme points of the breast. B. Landmarks were positioned tracking the inframammary fold. Landmarks were placed with at least one, every 30° with respect to the centre of the breast. If a clear infra-mammary fold was absent, a line of highest surface curvature was followed along the inferior aspect of the
EP
breast. The chest wall boundary was modelled by a flat plane, which was fitted to the
AC C
annotated points using least-squares fitting. C. Twelve points were annotated describing the “base of the breast” where the gland elevated from the chest wall as an ellipse. The ellipse was drawn, using a screen marker and desktop annotation program, first tracking the inframammary fold or a line of high curvature as surrogate using a stylus on a tablet PC. Then the less defined boundary segments were approximated by the elliptical shape constraint. Twelve points were subsequently placed every 30° along the ellipse. The chest wall was approximated using a flat plane fitted to the annotated points.
5
ACCEPTED MANUSCRIPT D. The same elliptical boundary annotation as C was used, but in contrast to A, B and C, the chest wall was not approximated by a flat plane. A curved surface described by thin-plate splines was used to model the chest wall19,31. The volume of the breast was calculated automatically based on the selected boundary definition
RI PT
from each 3D image using purpose-built software in Matlab32. Data evaluation
Patient demographics such as age, Body Mass Index (BMI) and bra size were documented and
SC
analysed, together with used implant sizes and mastectomy specimen weight. The relationship of excised tissue weight and implant chosen by the surgeon was explored as intraoperative specimen
M AN U
weight is sometimes seen as closely related to breast volume and can potentially serve as an approximate guide for reconstruction surgery33. Correlation was investigated by calculating the Spearman’s rank coefficient ρ. The coefficient of determination R2 34 was used to assess fit of a linear model based on weight to predict the chosen implants and the standard error comparing weight-based
TE D
implant selection and volume of the used implants, was calculated
The performance of 3D breast imaging for planning reconstructive surgery was evaluated, firstly investigating repeatability of the different boundary definitions (A, B, C, D) using the pre-surgery 3D
EP
images. The breast boundary was annotated five times for each patient using each of the four boundary definitions. Imaging position 5 (which covers the unaffected breast) was chosen for this
AC C
experiment as none of these images show mesh truncations, which can occur for the frontal Kinect position and potentially leaves the lateral aspect of very large breasts undescribed. In order to choose the most reliable boundary definition, Intra-Class Correlation (ICC)35 and the coefficient of repeatability36 were evaluated. The ICC relates the variance within repetitions ( ) with the variance
between subjects ( ): =
; repetitions for 95% of subjects are expected within one
. coefficient of repeatability: 2.77
The best imaging position was found, evaluating reproducibility with respect to how well patients adapted the same position in two sessions and the resulting influence on volume measurements. The
6
ACCEPTED MANUSCRIPT breast volume of the unaffected, healthy breast measured during both sessions was compared by calculating Spearman’s rank correlation and the coefficient of repeatability together with the 95% limits of agreement37.. 95% of the measurement discrepancies lie within the limits of agreement estimated as the mean discrepancy ± 1.96 standard deviations (of the discrepancies).
RI PT
Finally, power of preoperative 3D imaging to determine the optimal implant volume was assessed. The optimal implant volume ( ) was calculated by measuring present asymmetry after the reconstruction ( , −
! " #, ),
and correcting the used implant size (! ) by
SC
the offset: = ! + ( , −
! " #, )
M AN U
A linear model was used to describe the relationship between this optimal, corrected implant volume and the preoperative image-based volume measurement of the healthy breast. Predictive power was analysed by the coefficient of determination and compared with a model using specimen weight to predict optimal implant volume and a model taking into account both preoperatively measured breast
TE D
volume and specimen weight. Similarly, the relationship between the implant chosen by the experienced surgeon and the calculated implant size was assessed. The difference between predicted and optimal implant size was assessed in a leave one out fashion;
EP
i.e. the model was generated using all, but one, patients; implant size for the remaining patient was
AC C
predicted based on the model. This procedure was then repeated for all patients.
Results
Patient/ Intervention characteristics All patients (median age 51, range 34–61; median BMI 27, range 17–34; bra size range A–G cup) were treated with mastectomy for cancer (eight invasive carcinoma, two ductal cancer in situ). Seven patients in this study underwent skin-sparing mastectomy, three patients, skin and nipple sparing interventions. Eight patients were treated with immediate implant- based reconstruction, two patients
7
ACCEPTED MANUSCRIPT underwent a two-step procedure. All postoperative imaging was performed between 15 and 94 days after surgery, one patient was lost to follow-up. Implants used were silicone gel filled Mentor CPG™ 300 series implants, implant volumes used ranged from 155 to 620 ml (median 452 ml). Mastectomy specimens weighted between 165 and 988 g
RI PT
(median 396 g). Ease of use
The Kinect sensor was straight-forward to use in a clinical setting, even in small rooms. The
SC
acquisition protocol was simple and one imaging session with all seven 3D images took less than 30 minutes. Motion range in one patient was still limited after surgery and all poses with arms up were
M AN U
modified to be performed with arms elevated only within the pain-free motion range. The postoperative imaging, however, was, in all patients, used for result assessment only, not for implant size prediction.
Relationship between resection weight and used implant size
TE D
Weight of the excised specimen and chosen implant size were not significantly correlated (ρ=0.66, p=0.06). A coefficient of determination of R2=0.54 and a standard error of 200 ml comparing weight and implant size, illustrated the limited power of intraoperative weight measurement to predict
EP
implant selection (Figure 2).
AC C
Boundary annotation and definition of the chest wall For boundary annotation C, 95% of repeat volume measurements lay within 70 ml (= coefficient of repeatability). Variation among repeats was very small, compared to overall variation within the dataset (ICC = 0.996). Results were similar, but slightly worse for definition D, which used the same annotations as C, but approximated the chest wall as curved surface. Boundary definition A and B produced less stable results and were not used for following evaluations (coefficient of repeatability 167 ml for A,151 ml for B) (Table 2). Imaging Position
8
ACCEPTED MANUSCRIPT The most reproducible pose was achieved, when the patient sat upright with both arms above the head and hands resting on the head, the Kinect centrally positioned in front of the patient (95% of repeats within 141 ml and highest rank correlation comparing both imaging sessions ρ=0.93). The most variability was found for imaging position 2, where the patient was asked to sit upright and hold
RI PT
hands behind the back, which allowed some shoulder posture variation. Correlation between used implant volume and volume measured from the reconstructed breast was strongest for imaging position 6 and 7 (both ρ=0.78) (Table 3). Results for boundary annotation D were similar (coefficient
occurred for half of imaging positions (full results not shown).
M AN U
Implant choice
SC
of repeatability 110–307 ml) to method C, however, outliers, with maximum discrepancy >300 ml
When assessing the relationship between calculated optimal implant volume and the image-based volume of the healthy breast before the surgery, imaging position 5 produced the highest degree of correlation (ρ=0.88). Measurements taken from imaging position 4 showed the best fit of a linear model (R2=0.78) and most reliable prediction of implant size using the fitted model. Comparing
TE D
predicted implant sizes with the calculated ideal implant sizes resulted in a standard error of 120 ml. The weight and the choice of experienced surgeons resulted in similar, but lower coefficients of
(Figure 3).
EP
determination with the calculated optimal implant size (both: R2=0.74, standard error: 142/141 ml)
AC C
The patient with the largest breasts in our dataset exhibited the greatest discrepancy (37%) between selected implant size (620 ml) and optimal calculated implant size (982 ml for imaging position 4); a second patient showed >150 ml discrepancy, which equals an asymmetry of 100% in this case (Figure 4).
In order to access the quality of implant predictions using 3D imaging a leave-one-out validation was performed using images from image position 4: Data from 8 women was used to create a linear model, implant prediction based on this model was calculated for the 9th, which was left out for model creation. This was repeated for all datasets. In 61% of all configurations, discrepancy between optimal
9
ACCEPTED MANUSCRIPT implant volume and predicted implant volume was <10%; however, for 17% of configurations discrepancy was >30%.
RI PT
Discussion It was feasible to use the Kinect to measure preoperative breast volumes of cancer patients within less than 30 minutes. Using only one pose would reduce measurement time even further; we estimate less than 15 minutes are necessary including equipment setup.
SC
The most repeatable definition of the breast boundary was achieved, estimating the breast base as an ellipse (method C). This allowed indicating the breast boundary, where it is the most defined,
M AN U
including along the inframammary fold, and to extrapolate the boundary into segments where the breast is less clearly defined, such as superiorly, using the constraints of an ellipse. No clear benefits were found in modelling the internal boundary of the breast along the chest wall as a curved surface. This method produced several outliers, and may be less suited for clinical practice.
TE D
Potentially, body surface around the breast did not serve as a good indication of the shape of the chest wall, especially in patients with high BMI.
According to our measurements, it was most suitable to acquire 3D images of the patient’s breasts
EP
with implant prediction in mind, from 45° caudal angulation, either with the Kinect moving around the patient using KinectFusion (imaging position 4) or taking separate images for left and right breasts
AC C
from this position (imaging position 5/6). This is in agreement with the findings of Henseler and coworkers27, who suggested a tilted position of the patient instead of the imaging equipment. The patient should be seated with arms elevated resting on the head, in line with the finding that imaging position 3 was most reproducible. The use of a custom-build sensor holder for position 4 or 5/6 could potentially lead to similar reproducibility, eliminating uncertainties of using the Kinect in a hand-held fashion. Generally, reproducibility was likely to be underestimated in this study as some patients presented with visible weight change at the postoperative imaging session, which is not uncommon due to anxiety and/or commencing hormone therapy38.
10
ACCEPTED MANUSCRIPT Given the most suitable breast boundary definition and patient pose, we found preoperative imagebased volume measurement of the healthy breast indicated the optimal breast implant size with similar certainty to mastectomy specimen weight and to experienced surgeons. This technology could therefore aid planning for reconstruction surgery and implant procurement. Alternatively, the acquired
RI PT
3D models could enable the manufacture of customised implants, which is becoming possible with the advent of novel technologies such as 3D bio-printing 39–41.
SC
Conclusion
This study showed the feasibility of using fast, simple and inexpensive 3D imaging technology for
M AN U
predicting implant size before surgery, although there were significant technical challenges in determining breast volume by surface imaging.
A larger study is needed to further evaluate the potential of this technology to ensure breast cancer
Conflict of interest: None
EP
Ethical approval
TE D
patients get the optimum breast implants and best cosmetic outcome.
The study was approved by the NHS Research Ethics Committee Preston under the IRAS project
AC C
ID162918 and performed in accordance with the ethical standards of the 1964 Declaration of Helsinki and its subsequent amendments. The study was reported adhering to the STROBE guidelines. Acknowledgements
We would like to thank the nursing staff of the Nightingale Centre and the breast surgeons; Cliona Kirwan, James Harvey, John Murphy, Richard Johnson and Sumohan Chatterjee. We would also like to thank the charity Breast Cancer Now, which funded this study as part of a doctoral research studentship. The sponsors had no involvement in planning, conducting or publishing this study.
11
ACCEPTED MANUSCRIPT BIBILIOGAPHY
1.
Jeevan R, Cromwell D, Browne J, et al. National Mastectomy and Breast Reconstruction
2.
RI PT
Audit. NHS Inf. Cent. 2009. Tebbetts JB. A system for breast implant selection based on patient tissue characteristics and implant-soft tissue dynamics. Plast. Reconstr. Surg. 2002;109(4):1396-1409. doi:10.1097/00006534-200204010-00032.
Brown TP, Ringrose C, Hyland RE, Cole AA, Brotherston TM. A method of assessing female
SC
3.
doi:10.1054/bjps.1999.3110. 4.
M AN U
breast morphometry and its clinical application. Br. J. Plast. Surg. 1999;52(5):355-359.
Westreich M. Anthropomorphic breast measurement: protocol and results in 50 women with aesthetically perfect breasts and clinical application. Plast. Reconstr. Surg. 1997;100(2):46879. doi:10.1097/00006534-199708000-00032.
Hvilsom GB, Hölmich LR, Steding-Jessen M, et al. Delayed breast implant reconstruction: A
TE D
5.
10-year prospective study. J. Plast. Reconstr. Aesthetic Surg. 2011;64(11):1466-1474. doi:10.1016/j.bjps.2011.06.026.
Spear SL, Murphy DK, Slicton A, Walker PS. Inamed silicone breast implant core study
EP
6.
AC C
results at 6 years. Plast. Reconstr. Surg. 2007;120(7 Suppl 1):8S-16S; discussion 17S-18S. doi:10.1097/01.prs.0000286580.93214.df.
7.
Xi W, Perdanasari AT, Ong Y, et al. Objective breast volume, shape and surface area
assessment: A systematic review of breast measurement methods. Aesthetic Plast. Surg.
2014;38(6):1116-1130. doi:10.1007/s00266-014-0412-5. 8.
Kayar R, Civelek S, Cobanoglu M, Gungor O, Catal H, Emiroglu M. Five methods of breast volume measurement: a comparative study of measurements of specimen volume in 30 mastectomy cases. Breast cancer basic Clin. Res. 2011;5:43-52. doi:10.4137/BCBCR.S6128.
12
ACCEPTED MANUSCRIPT 9.
Choppin SB, Wheat JS, Gee M, Goyal A. The accuracy of breast volume measurement methods: A systematic review. The Breast 2016;28:121-129. doi:10.1016/j.breast.2016.05.010.
10.
Roostaeian J, Adams WP. Three-dimensional imaging for breast augmentation: Is this technology providing accurate simulations? Aesthetic Surg. J. 2014;34(6):857-875.
11.
RI PT
doi:10.1177/1090820X14538805. Mailey B, Freel A, Wong R, Pointer DT, Khoobehi K. Clinical accuracy and reproducibility of Portrait 3D surgical simulation platform in breast augmentation. Aesthetic Surg. J.
12.
SC
2013;33(1):84-92. doi:10.1177/1090820X12469807.
Donfrancesco A, Montemurro P, Hedén P. Three-dimensional simulated images in breast
M AN U
augmentation surgery. Plast. Reconstr. Surg. 2013;132(4):810-822. doi:10.1097/PRS.0b013e3182a014cb. 13.
Kovacs L, Eder M, Zimmermann A, et al. Three-dimensional evaluation of breast augmentation and the influence of anatomic and round implants on operative breast shape
14.
TE D
changes. Aesthetic Plast. Surg. 2012;36(4):879-887. doi:10.1007/s00266-012-9892-3. Gladilin E, Gabrielova B, Montemurro P, Hedén P. Customized planning of augmentation mammaplasty with silicon implants using three-dimensional optical body scans and
EP
biomechanical modeling of soft tissue outcome. Aesthetic Plast. Surg. 2011;35(4):494-501. doi:10.1007/s00266-010-9642-3. Eder M, Grabhorn A, Waldenfels F V., et al. Prediction of breast resection weight in reduction
AC C
15.
mammaplasty based on 3-dimensional surface imaging. Surg. Innov. 2013;20(4):356-364. doi:10.1177/1553350612460127.
16.
Szychta P, Raine C, Butterworth M, et al. Preoperative implant selection for two stage breast reconstruction with 3D imaging. Comput. Biol. Med. 2014;44:136-143. doi:10.1016/j.compbiomed.2013.09.013.
17.
Ahcan U, Bracun D, Zivec K, Pavlic R, Butala P. The use of 3D laser imaging and a new breast replica cast as a method to optimize autologous breast reconstruction after mastectomy.
13
ACCEPTED MANUSCRIPT The Breast 2012;21(2):183-189. doi:10.1016/j.breast.2011.09.009. 18.
Tomita K, Yano K, Hata Y, Nishibayashi A, Hosokawa K. DIEP flap breast reconstruction using 3-dimensional surface imaging and a printed mold. Plast. Reconstr. Surg. - Glob. Open 2015;3(3). doi:10.1097/GOX.0000000000000288. Tepper OM, Karp NS, Small K, et al. Three-dimensional imaging provides valuable clinical
RI PT
19.
data to aid in unilateral tissue expander-implant breast reconstruction. Breast J. 2008;14(6):543-50. doi:10.1111/j.1524-4741.2008.00645.x.
Tanabe YN, Honda T, Nakajima Y, Sakurai H, Nozaki M. Intraoperative application of three-
SC
20.
dimensional imaging for breast surgery. Scand. J. Plast. Reconstr. Surg. Hand Surg.
21.
M AN U
2005;39(6):349-352. doi:10.1080/02844310500296347.
Kovacs L, Eder M, Hollweck R, et al. Comparison between breast volume measurement using 3D surface imaging and classical techniques. The Breast 2007;16(2):137-45. doi:10.1016/j.breast.2006.08.001.
Catherwood T, McCaughan E, Greer E, Spence RAJ, McIntosh SA, Winder RJ. Validation of
TE D
22.
a passive stereophotogrammetry system for imaging of the breast: A geometric analysis. Med. Eng. Phys. 2011;33(8):900-905. doi:10.1016/j.medengphy.2011.02.005. Chae MP, Rozen WM, Spychal RT, Hunter-smith DJ. Breast volumetric analysis for aesthetic
EP
23.
planning in breast reconstruction: a literature review of techniques. Gland Surg.
24.
AC C
2015;5(6):212-226. doi:10.3978/j.issn.2227-684X.2015.10.03.
Oliveira HP, Silva MD, Magalhães A, Cardoso MJ, Cardoso JS. Is Kinect depth data accurate
for the aesthetic evaluation after breast cancer surgeries? In: Pattern Recognition and Image Analysis, LNCS7887. Springer Berlin Heidelberg; 2013:261-268.
25.
Lacher RM, Hipwell JH, Williams NR, Keshtgar MRS, Hawkes DJ, Stoyanov D. Low-cost surface reconstruction for aesthetic results assessment and prediction in breast cancer surgery. Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. EMBS 2015;2015-Novem:5871-5874. doi:10.1109/EMBC.2015.7319727.
14
ACCEPTED MANUSCRIPT 26.
Henseler H, Bonkat SK, Vogt PM, Rosenhahn B. The Kinect recording system for objective three- and four-dimensional breast assessment with image overlays. J. Plast. Reconstr. Aesthetic Surg. 2016;69(2):e27-e34. doi:10.1016/j.bjps.2015.10.021.
27.
Henseler H, Ju X, Ayoub A, Ray AK. The importance of the pose in three-dimensional
RI PT
imaging of the ptotic breast. J. Plast. Reconstr. Aesthetic Surg. 2013;66(11):1551-1556. doi:10.1016/j.bjps.2013.06.040. 28.
Reece GP, Merchant F, Andon J, et al. 3D surface imaging of the human female torso in
doi:10.1016/j.medengphy.2015.01.011.
Newcombe RA, Davison AJ, Izadi S, et al. KinectFusion: Real-time dense surface mapping
M AN U
29.
SC
upright to supine positions. Med. Eng. Phys. 2015;37(4):375-383.
and tracking. 10th IEEE Int. Symp. Mix. Augment. Real. 2011:127-136. doi:10.1109/ISMAR.2011.6092378. 30.
Qiao Q, Zhou G, Ling Y. Breast volume measurement in young Chinese women and clinical
31.
TE D
applications. Aesthetic Plast. Surg. 1997;21(5):362-368.
Kovacs L, Eder M, Hollweck R, et al. New aspects of breast volume measurement using 3dimensional surface imaging. Ann. Pastic Surg. 2006;57(6):602-10.
32.
EP
doi:10.1097/01.sap.0000235455.21775.6a.
Pöhlmann STL, Hewes J, Williamson AI, et al. Breast volume measurement using a games
AC C
console input device. In: Fujita H, Hara T, Muramatsu C, eds. IWDM 2014, LNCS 8539. Gifu: Springer International Publishing; 2014:666-673.
33.
Parmar C, West M, Pathak S, Nelson J, Martin L. Weight versus volume in breast surgery: an
observaParmar, C., West, M., Pathak, S., Nelson, J., & Martin, L. (2011). Weight versus volume in breast surgery: an observational study. JRSM Short Reports, 2(11), 87. http://doi.org/10.1258/shorts.2011.011070ti. JRSM Short Rep. 2011;2(11):87. doi:10.1258/shorts.2011.011070. 34.
Nagelkerke NJD. A note on a general definition of the coefficient of determination. Biometrika
15
ACCEPTED MANUSCRIPT 1991;78(3):691-692. doi:10.1093/biomet/78.3.691. 35.
Shrout PE, Fleiss JL. Intraclass correlations: Uses in assessing rater reliability. Psychol. Bull. 1979;86(2):420-428. doi:10.1037/0033-2909.86.2.420.
36.
Vaz S, Falkmer T, Passmore AE, Parsons R, Andreou P. The Case for Using the Repeatability
RI PT
Coefficient When Calculating Test-Retest Reliability. PLoS One 2013;8(9):1-7. doi:10.1371/journal.pone.0073990. 37.
Bland JM, Altman DG. Measuring agreement in method comparison studies. Stat. Methods
38.
SC
Med. Res. 1999;8(2):135-160. doi:10.1191/096228099673819272.
Vagenas D, DiSipio T, Battistutta D, et al. Weight and weight change following breast cancer:
M AN U
evidence from a prospective, population-based, breast cancer cohort study. BMC Cancer 2015;15(1):28. doi:10.1186/s12885-015-1026-2. 39.
Gerstle TL, Ibrahim AMS, Kim PS, Lee BT, Lin SJ. A Plastic Surgery Application in Evolution. Plast. Reconstr. Surg. 2014;133(2):446-451.
40.
TE D
doi:10.1097/01.prs.0000436844.92623.d3.
Tourassi GD, Vargas-Voracek R, Catarious DM, Floyd CE. Computer-assisted detection of mammographic masses: a template matching scheme based on mutual information. Med. Phys.
Rengier F, Mehndiratta A, Von Tengg-Kobligk H, et al. 3D printing based on imaging data:
AC C
41.
EP
2003;30:2123-2130. doi:10.1118/1.1589494.
Review of medical applications. Int. J. Comput. Assist. Radiol. Surg. 2010;5(4):335-341. doi:10.1007/s11548-010-0476-x.
16
ACCEPTED MANUSCRIPT
FIGURE LEGENDS Figure 1: Breast boundary definitions (A-D) applied to 3D image of a large breast (G-cup) with visible ptosis and pronounced inframammary fold and a small breast (A-cup) with absent ptosis, which made it
RI PT
necessary to estimate to inframammary fold, where curvature is highest
Figure 2: Relationship between resection weight and chosen implant volume (only nine points visible as data points overlap), the regression line is plotted as solid line, the line of equality is depicted as dashed
M AN U
SC
line
Figure 3: Relationship between image-based volume measurement, weight and implant choice by the surgeons with the calculated optimal implant volume (one patient was lost to follow-up, one dataset was incomplete lacking imaging position 4)
TE D
Figure 4: Patients with the largest discrepancy between used implant volume and calculated optimal implant volume, (a) largest volume discrepancy, (b) largest percent discrepancy; Kinect images pre- and post-surgery are shown; imaging position 03 allows judging asymmetry, for (b) a view from below is
AC C
EP
added, as asymmetry in the frontal view only mildly visible
17
ACCEPTED MANUSCRIPT
TABLES Table 1: The imaging protocol comprises seven imaging positions with different patient poses and Kinect locations 2
3
4
5
SC
Example
Patient
Sitting, arms
pose
arms back hanging
M AN U
Sitting, Sitting,
6
Sitting
Sitting
Sitting
arms above
arms
arms
head
above head
above head
arms up,
hands on
TE D
arms hanging
45° from
45° from
Moving in
the midline
the midline
On tripod,
On tripod,
an arc
and 45° of
and 45° of
Kinect
directly in
directly in
directly in
around the
cranial
cranial
location
front of
front of
front of
torso, 45°
angulation
angulation
subject
subject
of cranial
centred at
centred at
angulation
healthy
affected
breast
breast
EP
semi-
position,
On tripod,
AC C
Lying,
recumbent
head
subject
7
RI PT
1
45° of cranial angulation
18
ACCEPTED MANUSCRIPT
Table 2: Assessment of repeatability comparing sets of five repeat boundary annotations on each selected 3D image (most repeatable method shaded) Differences of repeats from patient mean [ml]
2
Max
correlation3
repeatability [ml]4
0.984
166.5
0.984
150.5
0.996
70.0
0.994
73.5
A
35.7
60.1
185.3
B
33.4
54.3
167.0
C
15.8
24.9
65.9
D
15.2
26.5
107.7
RI PT
Mean
SD
Absolute mean over all patients, for each patient the mean absolute difference was calculated, for the five individual
M AN U
1
Coefficient of
SC
1
Method
Intra-class Within subject
measurements with respect to the mean of those 5 repeats 2
) Calculated as square root of within-subject variance from a one-way ANOVA ('
3
Calculated as
4
Calculated as 2.77 ∙ '
()*
is the within-subject variance, ' is the between-groups variance from a one-way ANOVA , where '
AC C
EP
TE D
*
*
19
ACCEPTED MANUSCRIPT
Table 3: Assessment of reproducibility comparing volume of the healthy breast measured during the two imaging sessions for each imaging positions (best performance shaded) Spearman’s
Limits of Differences pre-post [ml]
Spearman’s
Coefficient
rank
of
rank
RI PT
agreement [ml]
correlation
correlation
Imaging Mean1
SD2
Max
lower
(post-
upper (pre-post)
position
repeatability 3
[ml]
implant)
108.6
269.1
-162.1
263.8
2
103.2
126.7
-266.8
-287.6
209.1
3
56.0
71.0
-152.4
-117.1
165.6
4
4.7
89.0
-122.4
-167.7
5/6
82.0
100.1
163.9
7
12.7
82.8
116.9
0.90
224.0
0.58
0.88
246.3
0.56
0.93
141.4
0.54
176.1
0.93
172.0
0.47
-195.3
197.3
0.91
183.5
0.78
-149.6
175.0
0.88
164.4
0.78
SC
80.5
M AN U
1
Absolute mean difference between pre and post-surgery imaging over all patients
2
Calculated from the signed difference (pre-post) over all patients
3
, where ' is the within-subject variance from a one-way ANOVA Calculated as 2.77 ∙ '
AC C
EP
TE D
1
20
ACCEPTED MANUSCRIPT
Table 4: Assessment of the relationship between optimal calculated implant volume and preoperative 3D imaging (best performance shaded) Spearman’s rank correlation (preoperative
Coefficient of determination (based on
position
volume measurement - optimal calculated
linear model predicting optimal implant
implant volume)
volume from preoperative volume
RI PT
Imaging
measurement)-
0.40
3
0.47
4
0.76
5/6
0.88
7
0.40
0.08
SC
2
0.35
0.42
M AN U
0.57
0.78
0.73 0.20
AC C
EP
TE D
1
21
AC C
EP
TE D
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT
1000 800 600 400
RI PT
Implant volume [cc]
1200
y = 0.341x + 252.47 R² = 0.5379
200 0
200
400
600
800
1000
1200
SC
0
AC C
EP
TE D
M AN U
Resection weight [g]
1200 1000 800 600
y = 1.1239x - 32.31 R² = 0.7773
400 200 0 0
200
400
600
800
1000
1200
SC
1000 800
y = 1.6415x - 282.7 R² = 0.7444
600 400 200 0 0
200
M AN U
1200
TE D
Calculated optimal implant volume [ml]
Implant volume derived form 3D imaging [ml]
400
600
800
1000
1200
Implants chosen by experienced surgeon [ml]
EP
1200 1000
AC C
Calculated optimal implant volume [ml]
RI PT
Calculated optimal implant volume [ml]
ACCEPTED MANUSCRIPT
800
y = 0.9567x + 125.69 R² = 0.7406
600 400 200 0
0
200
400
600
800
Specimen weight [g]
1000
1200
ACCEPTED MANUSCRIPT (a)
AC C
EP
TE D
M AN U
SC
RI PT
(b)