- Email: [email protected]
An automated intelligent diagnostic system for the interpretation of umbilical artery Doppler velocimetry
EUROPEAN JOURNALOF
RADIOLOGY ELSEVIER
European Journal of Radiology 23 (1996) 162-167
An automated intelligent diagnostic system for the interpretation of umbilical artery Doppler velocimetry M . S i n a n B e k s a ~ *a, A l i E g e m e n a, K u r t u l u ~ i z z e t o g l u b, G i i l ~ a h E r g i i n a, A y d a n M . E r k m e n b aBiomedical Engineering Unit, Department of Obstetrics and Gynecology, Hacettepe University, 06100 Ankara. Turkey bDepartment of Electrical and Electronics Engineering, Middle East Technical University, 06530 Ankara, Turkey
Accepted 4 June 1996
Abstract
The objective is to develop an automated intelligent diagnostic system for the interpretation of umbilical artery velocity waveforms. An ultrasound instrument with pulsed-wave Doppler is connected to a microcomputer by means of a frame grabber. After data acquisition, umbilical Doppler velocimetry is handled as a pattern recognition (feature extraction and classification) and decision-making problem. Automated image processing (enhancement, smoothing/thresholding and edge detection) and analysis are used for feature extraction. Six waveform indices obtained by feature extraction are used as input layer to vector quantization which classifies waveforms into six groups. A clinical decision is assigned to each group by the medical expert. Our system is trained by 278 and 380 waveform images of 94 normal and 157 high risk pregnancies, respectively. The system was tested with 193 and 61 images of normal and risky pregnancies; it was demonstrated that sensitivity and specificity of the system are 54.1% and 80.3%, respectively. Keywords: Ultrasound study, Doppler; Artery, umbilical; Pattern recognition
1. Introduction
Adequate blood flow through the umbilical circulation is essential to provide the fetus with oxygen and nutrients; therefore, umbilical cord haemodynamics should be a priority in perinatal surveillance [1-3]. There has been increasing interest in the application of Doppler ultrasound velocimetry as a fetal diagnostic tool [4-6]. However, there is serious debate on how and when to use umbilical artery Doppler velocimetry [7-10]. Pathophysiologic backgrounds of waveform changes are multifactorial and the biological rationales behind the conventional indices are unclear. These issues cause evaluation problems and the result is sophisticated, but with more subjective clinical interpretations. Recently, an automated intelligent diagnostic system for the interpretation of umbilical artery velocity waveforms has been developed to overcome * Corresponding author. Tel: +90 312 310 1011, +90 312 235 2828 (private); Fax: +90 312 310 5552.
problems arising from subjectivity of conventional interpretations [11,12]. In this study, umbilical artery velocity waveform interpretation is accepted as a pattern recognition and decision-making problem. Automated image processing and analysis is used to obtain waveform indices (feature extraction is the first step of pattern recognition), and vector quantization (second step) is used for the classification of waveforms [12,13]. 2. Materials and methods 2.1. Structure o f the system and data acquisition
An ultrasound instrument (General Electrics, Radius HR) with a 3.75 MHz probe (with pulsed wave Doppler) is connected to a 486 based microcomputer by means of a frame grabber (Iris DT 2853, Data Translation Inc., USA) and special acquisition software [14]. In this study, interpretation of flow velocity waveform (FVW) images is handled as a pattern recognition and
0720-048X/96/$15.00 © 1996 ElsevierScienceIreland Ltd. All fights reserved PII S0720-048X(96)01067-4
M.S. Beksa~ et al./ European Journal of Radiology 23 (1996) 162-167
163
decision-making problem. Borland Pascal version 7.0 is used. 2.2. Patwrn recognition and decision-making Pattern recognition is composed of feature extraction (calculation of Doppler waveform indices) and classification. Automated image processing and analysis are used for feature extraction [11]. Processing steps are enhancement, smoothing and thresholding, and edge detection (Fig. 1). In this study, we have selected 6 waveform indices to be calculated after image processing: A/B ratio, pulsatility index (PI), resistance index (RI), pulse related index (PRI), area ratio of wave (ARW) and angle between coincident slopes (ACS) (Fig. 2) [11,12]. Vector quantization is used for classification. Six waveform indices obtained by feature extraction make up the input layer of the generalized learning vector quantization (GLVQ) algorithm [13]. This algorithm classifies the velocity waveforms in different groups according to their specific configurations. The third step is clinical decision-making on the classified waveform groups. Clinical decisions are assigned to each waveform group by considering the group members through various 'evaluation variables'. Evaluative variables are maternal risk factors, interventions to the 'management/treatment' of risk factors, response during pregnancy, abnormal antepartum fetal heart rate testing results, intrauterine growth retardation, operative deliveries, umbilical cord acid-base measurements, APGAR score, admission into the special care baby unit, and a combination of these variables and perinatal mortality. 2.3. Clinical material
This study consisted of 278 and 380 umbilical artery velocity waveform images of 94 normal and 157 high risk pregnancies, respectively, for the training group. Table 1 gives the gestational risk factors of the study subjects. Recordings were made at 7 different gestational periods: 14-16, 17-20, 21-24, 25-28, 29-32, 33-36 and 37-42 gcstational weeks. The testing group consisted of 193 and 61 umbilical artery velocity waveform images of normal pregnancies and high risk pregnancies, respectively, with adverse outcome. Only subjects with hypertensive states of pregnancy, intrauterine growth retardation and hemolytic disease of the fetus (Rh[D] isoimmunization) were included in the testing group. All subjects were evaluated clinically and detailed "altrasonographic examinations were performed simultaneously. Doppler images were obtained from free floating umbilical cord and transferred to the computer environment to be processed.
Fig. 1. Imageprocessingand analysissteps. (a) Umbilicalartery velocity waveform image on computer monitor after data acquisition (enhancement). (b) The sameimageafter smoothingand thresholding. (c) Edgedetection and calculation of velocitywaveformindices.
164
M.S. Beksaf et al. /European Journal of Radiology 23 (1996) 162-167
1
Wave Length
Table 1 Pregnancy conditions and the risk factors of study subjects in the training data set Pregnancy and perinatal conditions
A
AJB i n d e x Pulsative index, PI Resistivity index, RI Pulse Related Index, PRI
=A / B = (A-B)/Mean Velocity = (A-B)tA = (A-B)/Wave Length
Normal pregnancies Hypertensive states of pregnancy Heart disease Maternal anemia Gestational diabetes Viral infection Preterm labour or PROM Intrauterine growth retardation Early or late pregnancy bleeding Hemolytic disease of fetus Intrauterine infection Genetic disorders Breech presentation Intrapartum fetal stress Neonatal interventions Others Total
j
No. of patients
No. of Doppler velocimetry
94 20 8 4 8 1 20 13 5 6 I 10 8 35 9 9
278 57 i3 4 18 4 59 28 12 17 I 24 20 77 27 19
251
658
specificity, false-positive rate ( F P R ) , false negative rate ( F N R ) and prevalence were used for the validation o f our computerized system. 3. R e s u l t s
Aw
O u r d a t a structure permits G L V Q to classify velocity waveforms into 6 groups. Fig. 3 shows the represen-
Area Ratio of Wave, ARW = A w / A b A n g l e Between C o i n c i d e n t Slopes, ACS = O
Fig. 2. Demonstration of six velocity waveform indices calculated by image processing and analysis.
All patients were evaluated b y the perinatal expert considering the following: the risk factors in pregnancy, pregnancy follow-up tools (obstetrical history, physical a n d gestational examinations, l a b o r a t o r y findings, etc.), the results o f perinatal surveillance tests (ultrasonography a n d a n t e p a r t u m fetal h e a r t rate testing), pathophysiological changes in l a b o u r a n d delivery, delivery modes, n e w b o r n examinations, a n d early/late neonatal interventions. P o o r o u t c o m e was defined according to these parameters: (1) mortality, (2) int r a p a r t u m fetal distress, (3) umbilical artery/vein acid-base measurements, (4) neonatal hypoxia/acidemia, (5) 5-min A P G A R scores _< 7 a n d m e c o n i u m aspiration, (6) intra-uterine growth r e t a r d a tion. Several statistical measures such as sensitivity,
Table 2 The phrases used to describe the classified groups by generalised learning vector quantization VQ group
Grading Medical expert interpretation
1
A
GOOD: velocity waveform (VW) represents sufficient blood supply; fetal condition is good.
2
B
FAIR: VW represents moderate blood supply; fetal condition is fair but there may be problems.
3
C
SATISFACTORY: VW represents intervened blood supply but blood supply is still satisfactory; there may be adverse outcome.
4
D
UNCERTAINTY: VW represents uncertainty; there is diffÉcultyin decision making: repeat the test or use other tests.
5
E
ADVERSE: VW represents reduced blood flow; there may be unstable fetal condition.
6
F
CRITICAL: VW represents insufficient blood supply; there may be CRITICAL fetal condition.
M.S. Beksa~ et aL /European Journal of Radiology 23 (1996) 162-167
1
165
2 i
I
3
..:] :
....
::
4
:i~ .:.:~ ..~i-. •::~ : :: ~:
~ i
..................................................
• .!., -:,
5
6
Fig. 3. Vector quantization classifies velocity waveforms of each gestational period in 6 groups. The representative waveforms of 6 groups for 29-32 gestational weeks.
M.S. Beksa~ et al./European Journal of Radiology 23 (1996) 162-167
166
Table 3 Confusion matrix comparing similarities and dissimilarities between
the medical expert and the analysis program (testing group) Medical expert
Analysis program Normal
Adverse
Normal
155
38
Adverse
28
33
tative waveforms of each group for the 4th (29-32 gestational weeks)gestational period. These 6 groups are the output layer of the GLVQ. Each group of all gestational periods is described from the 'clinical decision-making' point of view considering the member patients and conventional approaches (evaluation of representative waveform of each group by means of classical indices). This stage is the only stage supervised by the medical expert. The predictive value of the system is highly dependent on the knowledge and experience of the perinatologist and also on the type of the data set used for training. On the other hand, the trainable character of the system permits each medical center to create its own version. Table 2 gives the phrases used by our medical expert to describe different groups (these phrases may differ between different medical centers). The system (Bolu System version 2.0) is tested with 193 and 61 umbilical artery velocity waveforms of normal and high risk pregnancies, respectively, of different gestational periods. Table 3 gives the confusion matrix comparing similarities and dissimilarities between the medical expert and the analysis program. The first three groups (A, good; B, fair; C, satisfactory) are accepted as normal, while the last two groups (E, adverse; F, critical) are defined as abnormal. 'Group D' (uncertainty) is excluded and the test is repeated within 1 h until a definite result is obtained. Table 4 gives the predictive values of the analysis program. The sensitivity and specificity of the system proved to be 54.1% and 80.3%, respectively. 4. Discussion
Umbilical artery velocity waveforms have been studied extensively over the past decade [15-18]. Various Doppler waveform indices were described relying on the pulsatility, which is the difference between the peak systolic and end-diastolic components of the maxiTable 4 Predictive values of the system (Version 2.0) Positive predictive value Negative predictive value False positive rate False negative rate Prevalence
46.5% 84.7% 19.7% 45-9% 28.2%
mum frequency shift, and the end-diastolic component itself [19-22]. Several workers described comprehensive approaches to the Doppler waveform analysis, but there is no current evidence that these approaches are superior to the classical Doppler indices [3]. The main problem is the lack of information related to what these indices really mean, and it has generally been assumed that changes in indices are due to alterations in the resistance downstream from the point of measurement. It is, however, well known from physiology that velocity waveform changes are also influenced by a number of other factors [17]. Although there is no clear evidence on the predictive value of conventionally-used classical indices, various studies have reported on the clinical significance of umbilical Doppler velocimetry in different clinical conditions [23,24]. Doppler measurements of umbilical and fetal circulations and the relationship with umbilical cord acid-base status were also demonstrated [18,251. The results of umbilical Doppler velocimetry studies are not promising and the place of this methodology with its current form and status should be reconsidered. Lack of clear evidence on its predictive value necessitates application of other surveillance methodologies and consideration of other fetal arteries during the course of measurements [8,9,26]. The first issue to be reconsidered is the selection of indices and rearrangement of the way these indices are used in clinical decision-making. Once the physiologic and clinical information behind velocity waveforms is clarified, it will then be possible to design proper clinical trials. Until then, it is better to handle waveform indices as nonspecific parameters and use them as an input layer to mathematical classifiers which are able to differentiate waveforms according to pattern specifications. Recently, an intelligent diagnostic system using supervised neural networks has been described for the evaluation of umbilical velocity waveforms [12]. A backpropagation learning algorithm (supervised neural network) is used for the classification of waveforms by using the mean values of 5 indices as input parameters [12]. In that system, medical expert know-how was essential in arranging input data. In the present study, we handled umbilical waveform analysis as a pattern recognition and decision-making problem. Automated image processing and analysis have been used to obtain standardization and 6 indices (A/B ratio, PI, RI, PRI, ARW and ACS) are extracted from velocity waveforms (feature extraction). Extracted features of all patients, without considering the clinical conditions and risk factors, are used as input layer to GLVQ which classifies the waveforms into 6 groups. Waveforms of each monthly gestational period are classified separately and at this point the medical expert is asked to describe clinical conditions of members of each group. A clinical decision is assigned to each group considering both the classical information through waveform indices and the
M.S. Beksa~ et al. /European Journal of Radiology 23 (1996) 162-167
risk factors of members of the groups. Unlike previously described intelligent diagnostic systems with neural networks, in this study medical expert intervention comes after classification. We believe that such an approach will bring more objectivity to the medical interpretation of umbilical velocity waveforms. This system uses various waveform indices, including the classical ones, in an unsupervised manner because we believe that clinical information behind them is unclear. It is important for analysis program to be sensitive in identifying infants at risk for hypoxic/acidemic morbidity via an abnormal test. On the other hand, it is important that a normal fetus be identified by a highly specific test to avoid unwarranted intervention. A laboratory test is commonly assessed by looking at statistical measures of its predictability of outcome. It has been reported that optimum validative criteria to be satisfied on the systems for the interpretation of any screening test should include: (1) sensitivity, 50%; (2) specificity, 94%; (3) positive predictive value (PPV), 50%; (4) negative predictive value (NPV), 94%; all based on an application of Bayes' theorem, assuming a disease prevalence of 10% [271. The predictive value of this system is promising and comparable with other methodologies [27]. We believe that objective approaches in Doppler velocimetry will stimulate advances in this field; specifically, further investigations are necessary to improve the predictive value of the system. The sensitivity and specificity of the system are 54.1% and 80.3%, respectively. The false positive rate is 19.7% and this rate is due to the high incidence of risk factors in the study population (prevalence 28.2%). Using different and well-defined training sets and improvements in the learning algorithm may improve these statistical measures. References [I] Fitzgerald DE, D r u m JE. Non-invasive measurement of human fetal circulation using ultrasound: a new method. Br Med J 1977; 2: 1450. [2] Mori A, Iwashita M, Takeda Y. Haemodynamic changes in IUGR fetus with chronic hypoxia evaluated by fetal heart rate monitoring and Doppler measurement of blood flow velocity. Med Biol Eng Comput 1993; Suppl 31: 549-558. [3] Maulik D, Yarlagadda P, Downing G. Doppler velocimetry in obstetrics. Obstet Gynecol Clin North Am 1990; 17: 163-186. [4] Newnham JP, O'Dea MR, Reid KP, Diepeveen DA. Doppler flow velocity waveform analysis in high risk pregnancies: a randomized controlled trial. Br J Obstet Gynaecoi 1991; 98; 956-963. [5] Bruner JB, Levy DW, Arger PH. Doppler ultrasonography of the umbilical cord in complicated pregnancies. South Med J 1993; 86: 418-422. [6] Omtzigt AM, Reuwer PJ, Bruinse HW. A randomized controlled trial on the clinical value of umbilical Doppler velocimetry in antenatal care. Am J Obstet Gynecol 1994; 170: 625-634. [7] vanVugt JMG. Validity of umbilical artery blood velocimetry in the prediction of intrauterine growth retardation and fetal compromise. J Perinat Med 1991; 19: 15-20. 18] Anteby EY, Tadmar O, Revel A, Yagel S. Post-term pregnan-
167
cies with normal cardiotocographs and amniotic fluid columns: the role of Doppler evaluation in predicting perinatai outcome. Eur J Obstet Gynecol Reprod Biol 1994; 54: 93-98. [9] Chan FY, Lain C, Lain YH, To WK, Pun TC, Lee CP. Umbilical artery Doppler velocimetry compared with fetal heart rate monitoring as a labor admission test. Eur J Obstet Gynecol Reprod Biol 1994; 54: 1-6. [10] Pattinson RC, Norman K, Odendaal HJ. The use of Doppler veiocimetry of the umbilical artery before 24 weeks gestation to screen for high-risk pregnancies. S Aft Med J 1993; 83: 734-736. [11] Ba~ran F, Beksa~; MS, Erkmen AM. An automated system for the evaluation of antepartum blood flow velocity waveforms: feature extraction. Turk J Med Sci 1993; 18: 227-232. [12] Beksa~ MS, Ba~aran F, Eskiizmirliler S, Erkmen A, Y6riikan S. A computerized diagnostic system for the interpretation of umbilical artery blood flow velocity waveforms. Eur J Obstet Gynecol Reprod Biol 1996; 64: 37-42. [13] Linde Y, Buzo A, Gray RM. An algorithm for vector quantizer design. IEEE Trans Commun 1980; 28: 84. [14] Beksa~ MS, G6ren C, Kartal G, Erkmen AM. An image processing for umbilical artery blood flow velocity waveforms. Turk J Med Res 1992; 10: 298-300. [15] Trudinger BJ, Cook CM, Giles WB, Connely A, Thompson RS. Umbilical artery flow velocity waveforms in high risk pregnancies. Lancet 1987; 24: 188-190. [16] Erkine RLA, Ritchie JWK. Umbilical artery blood flow characteristics in normal and growth-retarded fetuses. Br J Obstet Gynecol 1985; 92: 605-610. [17] deHaan J. Pathophysiologic backgrounds of changes in Doppler flow indices. In: van Geijn HP, Copray FJA, editors. A critical appraisal of fetal surveillance. Amsterdam: Elsevier Science BV, 1994; 470-482. [18] Arduini D, Rizzo G, Romanini C, Mancuso S. Are blood flow velocity waveforms related to umbilical cord acid-base status in the human fetus? Gynecol Obstet Invest 1989; 27: 183-187. [19] Pourcelot L. Applications clinique de i'examen Doppler transcutane. In: Pourcelot L, editor. Velocimetric Ultrasonore Doppler. INSERM 1974; 34: 213. [20] Gosling RG, King DH. Ultrasound angiology. In: Marcus AW, Adamson J, editors. Arteries and Veins. Edinburgh, UK: Churchill-Livingstone, 1975; 61. [21] Fairle FM, Moretti M, Walker J J, Sibai BM. Determinants of perinatal outcome in pregnancy-induced hypertension with absence of umbilical artery end-diastolic frequencies. Am J Obstet Gynecol 1991; 164: 1084-1089. [22] Rochelson B, Fleischer A, Penny B, Winter D. The significance of absent end-diastolic velocity in umbilical artery velocity waveforms. Am J Obstet Gynecol 1987; 156: 1213-1218. [23] Fleischer A, Schulman H, Farmakides G, Bracero L, Blattner P, Randolph G. Umbilical artery velocity waveforms and intrauterine growth retardation. Am J Obstet Gynecol 1985; 151: 502-5. [24] Brar HS, Medearis AL, DeVare GR, Platt LD. Maternal and fetal blood flow velocity waveforms in patients with preterm labor: prediction of successful tocolysis. Am J Obstet Gynecol 1988; 159: 947-50. [25] Bilardo CM, Nicolaldes KH, Campbell S. Doppler measurements of fetal and uteroplacental circulations: relationship with umbilical venous blood gases measured at cordocentesis. Am J Obstet Gynecol 1987; 69: 705-9. [26] Wladimiroff JW, Wijngaard AGW, Degani S, Noordam M J, van Eyck J, Tonge HM. Cerebral and umbilical arterial blood flow velocity waveforms in normal and growth-retarded pregnancies. Obstet Oynecol 1987; 69: 705-9. [271 Beksa¢ MS, Onderoolu LS, Ozdemir K, Karaka~ O. The validation of a computerized system for the interpretation of the antepartum fetal heart rate tracings (version 89/2.34). Eur J Obstet Gynecol Reprod Biol 1991; 42: 9-14.
RADIOLOGY ELSEVIER
European Journal of Radiology 23 (1996) 162-167
An automated intelligent diagnostic system for the interpretation of umbilical artery Doppler velocimetry M . S i n a n B e k s a ~ *a, A l i E g e m e n a, K u r t u l u ~ i z z e t o g l u b, G i i l ~ a h E r g i i n a, A y d a n M . E r k m e n b aBiomedical Engineering Unit, Department of Obstetrics and Gynecology, Hacettepe University, 06100 Ankara. Turkey bDepartment of Electrical and Electronics Engineering, Middle East Technical University, 06530 Ankara, Turkey
Accepted 4 June 1996
Abstract
The objective is to develop an automated intelligent diagnostic system for the interpretation of umbilical artery velocity waveforms. An ultrasound instrument with pulsed-wave Doppler is connected to a microcomputer by means of a frame grabber. After data acquisition, umbilical Doppler velocimetry is handled as a pattern recognition (feature extraction and classification) and decision-making problem. Automated image processing (enhancement, smoothing/thresholding and edge detection) and analysis are used for feature extraction. Six waveform indices obtained by feature extraction are used as input layer to vector quantization which classifies waveforms into six groups. A clinical decision is assigned to each group by the medical expert. Our system is trained by 278 and 380 waveform images of 94 normal and 157 high risk pregnancies, respectively. The system was tested with 193 and 61 images of normal and risky pregnancies; it was demonstrated that sensitivity and specificity of the system are 54.1% and 80.3%, respectively. Keywords: Ultrasound study, Doppler; Artery, umbilical; Pattern recognition
1. Introduction
Adequate blood flow through the umbilical circulation is essential to provide the fetus with oxygen and nutrients; therefore, umbilical cord haemodynamics should be a priority in perinatal surveillance [1-3]. There has been increasing interest in the application of Doppler ultrasound velocimetry as a fetal diagnostic tool [4-6]. However, there is serious debate on how and when to use umbilical artery Doppler velocimetry [7-10]. Pathophysiologic backgrounds of waveform changes are multifactorial and the biological rationales behind the conventional indices are unclear. These issues cause evaluation problems and the result is sophisticated, but with more subjective clinical interpretations. Recently, an automated intelligent diagnostic system for the interpretation of umbilical artery velocity waveforms has been developed to overcome * Corresponding author. Tel: +90 312 310 1011, +90 312 235 2828 (private); Fax: +90 312 310 5552.
problems arising from subjectivity of conventional interpretations [11,12]. In this study, umbilical artery velocity waveform interpretation is accepted as a pattern recognition and decision-making problem. Automated image processing and analysis is used to obtain waveform indices (feature extraction is the first step of pattern recognition), and vector quantization (second step) is used for the classification of waveforms [12,13]. 2. Materials and methods 2.1. Structure o f the system and data acquisition
An ultrasound instrument (General Electrics, Radius HR) with a 3.75 MHz probe (with pulsed wave Doppler) is connected to a 486 based microcomputer by means of a frame grabber (Iris DT 2853, Data Translation Inc., USA) and special acquisition software [14]. In this study, interpretation of flow velocity waveform (FVW) images is handled as a pattern recognition and
0720-048X/96/$15.00 © 1996 ElsevierScienceIreland Ltd. All fights reserved PII S0720-048X(96)01067-4
M.S. Beksa~ et al./ European Journal of Radiology 23 (1996) 162-167
163
decision-making problem. Borland Pascal version 7.0 is used. 2.2. Patwrn recognition and decision-making Pattern recognition is composed of feature extraction (calculation of Doppler waveform indices) and classification. Automated image processing and analysis are used for feature extraction [11]. Processing steps are enhancement, smoothing and thresholding, and edge detection (Fig. 1). In this study, we have selected 6 waveform indices to be calculated after image processing: A/B ratio, pulsatility index (PI), resistance index (RI), pulse related index (PRI), area ratio of wave (ARW) and angle between coincident slopes (ACS) (Fig. 2) [11,12]. Vector quantization is used for classification. Six waveform indices obtained by feature extraction make up the input layer of the generalized learning vector quantization (GLVQ) algorithm [13]. This algorithm classifies the velocity waveforms in different groups according to their specific configurations. The third step is clinical decision-making on the classified waveform groups. Clinical decisions are assigned to each waveform group by considering the group members through various 'evaluation variables'. Evaluative variables are maternal risk factors, interventions to the 'management/treatment' of risk factors, response during pregnancy, abnormal antepartum fetal heart rate testing results, intrauterine growth retardation, operative deliveries, umbilical cord acid-base measurements, APGAR score, admission into the special care baby unit, and a combination of these variables and perinatal mortality. 2.3. Clinical material
This study consisted of 278 and 380 umbilical artery velocity waveform images of 94 normal and 157 high risk pregnancies, respectively, for the training group. Table 1 gives the gestational risk factors of the study subjects. Recordings were made at 7 different gestational periods: 14-16, 17-20, 21-24, 25-28, 29-32, 33-36 and 37-42 gcstational weeks. The testing group consisted of 193 and 61 umbilical artery velocity waveform images of normal pregnancies and high risk pregnancies, respectively, with adverse outcome. Only subjects with hypertensive states of pregnancy, intrauterine growth retardation and hemolytic disease of the fetus (Rh[D] isoimmunization) were included in the testing group. All subjects were evaluated clinically and detailed "altrasonographic examinations were performed simultaneously. Doppler images were obtained from free floating umbilical cord and transferred to the computer environment to be processed.
Fig. 1. Imageprocessingand analysissteps. (a) Umbilicalartery velocity waveform image on computer monitor after data acquisition (enhancement). (b) The sameimageafter smoothingand thresholding. (c) Edgedetection and calculation of velocitywaveformindices.
164
M.S. Beksaf et al. /European Journal of Radiology 23 (1996) 162-167
1
Wave Length
Table 1 Pregnancy conditions and the risk factors of study subjects in the training data set Pregnancy and perinatal conditions
A
AJB i n d e x Pulsative index, PI Resistivity index, RI Pulse Related Index, PRI
=A / B = (A-B)/Mean Velocity = (A-B)tA = (A-B)/Wave Length
Normal pregnancies Hypertensive states of pregnancy Heart disease Maternal anemia Gestational diabetes Viral infection Preterm labour or PROM Intrauterine growth retardation Early or late pregnancy bleeding Hemolytic disease of fetus Intrauterine infection Genetic disorders Breech presentation Intrapartum fetal stress Neonatal interventions Others Total
j
No. of patients
No. of Doppler velocimetry
94 20 8 4 8 1 20 13 5 6 I 10 8 35 9 9
278 57 i3 4 18 4 59 28 12 17 I 24 20 77 27 19
251
658
specificity, false-positive rate ( F P R ) , false negative rate ( F N R ) and prevalence were used for the validation o f our computerized system. 3. R e s u l t s
Aw
O u r d a t a structure permits G L V Q to classify velocity waveforms into 6 groups. Fig. 3 shows the represen-
Area Ratio of Wave, ARW = A w / A b A n g l e Between C o i n c i d e n t Slopes, ACS = O
Fig. 2. Demonstration of six velocity waveform indices calculated by image processing and analysis.
All patients were evaluated b y the perinatal expert considering the following: the risk factors in pregnancy, pregnancy follow-up tools (obstetrical history, physical a n d gestational examinations, l a b o r a t o r y findings, etc.), the results o f perinatal surveillance tests (ultrasonography a n d a n t e p a r t u m fetal h e a r t rate testing), pathophysiological changes in l a b o u r a n d delivery, delivery modes, n e w b o r n examinations, a n d early/late neonatal interventions. P o o r o u t c o m e was defined according to these parameters: (1) mortality, (2) int r a p a r t u m fetal distress, (3) umbilical artery/vein acid-base measurements, (4) neonatal hypoxia/acidemia, (5) 5-min A P G A R scores _< 7 a n d m e c o n i u m aspiration, (6) intra-uterine growth r e t a r d a tion. Several statistical measures such as sensitivity,
Table 2 The phrases used to describe the classified groups by generalised learning vector quantization VQ group
Grading Medical expert interpretation
1
A
GOOD: velocity waveform (VW) represents sufficient blood supply; fetal condition is good.
2
B
FAIR: VW represents moderate blood supply; fetal condition is fair but there may be problems.
3
C
SATISFACTORY: VW represents intervened blood supply but blood supply is still satisfactory; there may be adverse outcome.
4
D
UNCERTAINTY: VW represents uncertainty; there is diffÉcultyin decision making: repeat the test or use other tests.
5
E
ADVERSE: VW represents reduced blood flow; there may be unstable fetal condition.
6
F
CRITICAL: VW represents insufficient blood supply; there may be CRITICAL fetal condition.
M.S. Beksa~ et aL /European Journal of Radiology 23 (1996) 162-167
1
165
2 i
I
3
..:] :
....
::
4
:i~ .:.:~ ..~i-. •::~ : :: ~:
~ i
..................................................
• .!., -:,
5
6
Fig. 3. Vector quantization classifies velocity waveforms of each gestational period in 6 groups. The representative waveforms of 6 groups for 29-32 gestational weeks.
M.S. Beksa~ et al./European Journal of Radiology 23 (1996) 162-167
166
Table 3 Confusion matrix comparing similarities and dissimilarities between
the medical expert and the analysis program (testing group) Medical expert
Analysis program Normal
Adverse
Normal
155
38
Adverse
28
33
tative waveforms of each group for the 4th (29-32 gestational weeks)gestational period. These 6 groups are the output layer of the GLVQ. Each group of all gestational periods is described from the 'clinical decision-making' point of view considering the member patients and conventional approaches (evaluation of representative waveform of each group by means of classical indices). This stage is the only stage supervised by the medical expert. The predictive value of the system is highly dependent on the knowledge and experience of the perinatologist and also on the type of the data set used for training. On the other hand, the trainable character of the system permits each medical center to create its own version. Table 2 gives the phrases used by our medical expert to describe different groups (these phrases may differ between different medical centers). The system (Bolu System version 2.0) is tested with 193 and 61 umbilical artery velocity waveforms of normal and high risk pregnancies, respectively, of different gestational periods. Table 3 gives the confusion matrix comparing similarities and dissimilarities between the medical expert and the analysis program. The first three groups (A, good; B, fair; C, satisfactory) are accepted as normal, while the last two groups (E, adverse; F, critical) are defined as abnormal. 'Group D' (uncertainty) is excluded and the test is repeated within 1 h until a definite result is obtained. Table 4 gives the predictive values of the analysis program. The sensitivity and specificity of the system proved to be 54.1% and 80.3%, respectively. 4. Discussion
Umbilical artery velocity waveforms have been studied extensively over the past decade [15-18]. Various Doppler waveform indices were described relying on the pulsatility, which is the difference between the peak systolic and end-diastolic components of the maxiTable 4 Predictive values of the system (Version 2.0) Positive predictive value Negative predictive value False positive rate False negative rate Prevalence
46.5% 84.7% 19.7% 45-9% 28.2%
mum frequency shift, and the end-diastolic component itself [19-22]. Several workers described comprehensive approaches to the Doppler waveform analysis, but there is no current evidence that these approaches are superior to the classical Doppler indices [3]. The main problem is the lack of information related to what these indices really mean, and it has generally been assumed that changes in indices are due to alterations in the resistance downstream from the point of measurement. It is, however, well known from physiology that velocity waveform changes are also influenced by a number of other factors [17]. Although there is no clear evidence on the predictive value of conventionally-used classical indices, various studies have reported on the clinical significance of umbilical Doppler velocimetry in different clinical conditions [23,24]. Doppler measurements of umbilical and fetal circulations and the relationship with umbilical cord acid-base status were also demonstrated [18,251. The results of umbilical Doppler velocimetry studies are not promising and the place of this methodology with its current form and status should be reconsidered. Lack of clear evidence on its predictive value necessitates application of other surveillance methodologies and consideration of other fetal arteries during the course of measurements [8,9,26]. The first issue to be reconsidered is the selection of indices and rearrangement of the way these indices are used in clinical decision-making. Once the physiologic and clinical information behind velocity waveforms is clarified, it will then be possible to design proper clinical trials. Until then, it is better to handle waveform indices as nonspecific parameters and use them as an input layer to mathematical classifiers which are able to differentiate waveforms according to pattern specifications. Recently, an intelligent diagnostic system using supervised neural networks has been described for the evaluation of umbilical velocity waveforms [12]. A backpropagation learning algorithm (supervised neural network) is used for the classification of waveforms by using the mean values of 5 indices as input parameters [12]. In that system, medical expert know-how was essential in arranging input data. In the present study, we handled umbilical waveform analysis as a pattern recognition and decision-making problem. Automated image processing and analysis have been used to obtain standardization and 6 indices (A/B ratio, PI, RI, PRI, ARW and ACS) are extracted from velocity waveforms (feature extraction). Extracted features of all patients, without considering the clinical conditions and risk factors, are used as input layer to GLVQ which classifies the waveforms into 6 groups. Waveforms of each monthly gestational period are classified separately and at this point the medical expert is asked to describe clinical conditions of members of each group. A clinical decision is assigned to each group considering both the classical information through waveform indices and the
M.S. Beksa~ et al. /European Journal of Radiology 23 (1996) 162-167
risk factors of members of the groups. Unlike previously described intelligent diagnostic systems with neural networks, in this study medical expert intervention comes after classification. We believe that such an approach will bring more objectivity to the medical interpretation of umbilical velocity waveforms. This system uses various waveform indices, including the classical ones, in an unsupervised manner because we believe that clinical information behind them is unclear. It is important for analysis program to be sensitive in identifying infants at risk for hypoxic/acidemic morbidity via an abnormal test. On the other hand, it is important that a normal fetus be identified by a highly specific test to avoid unwarranted intervention. A laboratory test is commonly assessed by looking at statistical measures of its predictability of outcome. It has been reported that optimum validative criteria to be satisfied on the systems for the interpretation of any screening test should include: (1) sensitivity, 50%; (2) specificity, 94%; (3) positive predictive value (PPV), 50%; (4) negative predictive value (NPV), 94%; all based on an application of Bayes' theorem, assuming a disease prevalence of 10% [271. The predictive value of this system is promising and comparable with other methodologies [27]. We believe that objective approaches in Doppler velocimetry will stimulate advances in this field; specifically, further investigations are necessary to improve the predictive value of the system. The sensitivity and specificity of the system are 54.1% and 80.3%, respectively. The false positive rate is 19.7% and this rate is due to the high incidence of risk factors in the study population (prevalence 28.2%). Using different and well-defined training sets and improvements in the learning algorithm may improve these statistical measures. References [I] Fitzgerald DE, D r u m JE. Non-invasive measurement of human fetal circulation using ultrasound: a new method. Br Med J 1977; 2: 1450. [2] Mori A, Iwashita M, Takeda Y. Haemodynamic changes in IUGR fetus with chronic hypoxia evaluated by fetal heart rate monitoring and Doppler measurement of blood flow velocity. Med Biol Eng Comput 1993; Suppl 31: 549-558. [3] Maulik D, Yarlagadda P, Downing G. Doppler velocimetry in obstetrics. Obstet Gynecol Clin North Am 1990; 17: 163-186. [4] Newnham JP, O'Dea MR, Reid KP, Diepeveen DA. Doppler flow velocity waveform analysis in high risk pregnancies: a randomized controlled trial. Br J Obstet Gynaecoi 1991; 98; 956-963. [5] Bruner JB, Levy DW, Arger PH. Doppler ultrasonography of the umbilical cord in complicated pregnancies. South Med J 1993; 86: 418-422. [6] Omtzigt AM, Reuwer PJ, Bruinse HW. A randomized controlled trial on the clinical value of umbilical Doppler velocimetry in antenatal care. Am J Obstet Gynecol 1994; 170: 625-634. [7] vanVugt JMG. Validity of umbilical artery blood velocimetry in the prediction of intrauterine growth retardation and fetal compromise. J Perinat Med 1991; 19: 15-20. 18] Anteby EY, Tadmar O, Revel A, Yagel S. Post-term pregnan-
167
cies with normal cardiotocographs and amniotic fluid columns: the role of Doppler evaluation in predicting perinatai outcome. Eur J Obstet Gynecol Reprod Biol 1994; 54: 93-98. [9] Chan FY, Lain C, Lain YH, To WK, Pun TC, Lee CP. Umbilical artery Doppler velocimetry compared with fetal heart rate monitoring as a labor admission test. Eur J Obstet Gynecol Reprod Biol 1994; 54: 1-6. [10] Pattinson RC, Norman K, Odendaal HJ. The use of Doppler veiocimetry of the umbilical artery before 24 weeks gestation to screen for high-risk pregnancies. S Aft Med J 1993; 83: 734-736. [11] Ba~ran F, Beksa~; MS, Erkmen AM. An automated system for the evaluation of antepartum blood flow velocity waveforms: feature extraction. Turk J Med Sci 1993; 18: 227-232. [12] Beksa~ MS, Ba~aran F, Eskiizmirliler S, Erkmen A, Y6riikan S. A computerized diagnostic system for the interpretation of umbilical artery blood flow velocity waveforms. Eur J Obstet Gynecol Reprod Biol 1996; 64: 37-42. [13] Linde Y, Buzo A, Gray RM. An algorithm for vector quantizer design. IEEE Trans Commun 1980; 28: 84. [14] Beksa~ MS, G6ren C, Kartal G, Erkmen AM. An image processing for umbilical artery blood flow velocity waveforms. Turk J Med Res 1992; 10: 298-300. [15] Trudinger BJ, Cook CM, Giles WB, Connely A, Thompson RS. Umbilical artery flow velocity waveforms in high risk pregnancies. Lancet 1987; 24: 188-190. [16] Erkine RLA, Ritchie JWK. Umbilical artery blood flow characteristics in normal and growth-retarded fetuses. Br J Obstet Gynecol 1985; 92: 605-610. [17] deHaan J. Pathophysiologic backgrounds of changes in Doppler flow indices. In: van Geijn HP, Copray FJA, editors. A critical appraisal of fetal surveillance. Amsterdam: Elsevier Science BV, 1994; 470-482. [18] Arduini D, Rizzo G, Romanini C, Mancuso S. Are blood flow velocity waveforms related to umbilical cord acid-base status in the human fetus? Gynecol Obstet Invest 1989; 27: 183-187. [19] Pourcelot L. Applications clinique de i'examen Doppler transcutane. In: Pourcelot L, editor. Velocimetric Ultrasonore Doppler. INSERM 1974; 34: 213. [20] Gosling RG, King DH. Ultrasound angiology. In: Marcus AW, Adamson J, editors. Arteries and Veins. Edinburgh, UK: Churchill-Livingstone, 1975; 61. [21] Fairle FM, Moretti M, Walker J J, Sibai BM. Determinants of perinatal outcome in pregnancy-induced hypertension with absence of umbilical artery end-diastolic frequencies. Am J Obstet Gynecol 1991; 164: 1084-1089. [22] Rochelson B, Fleischer A, Penny B, Winter D. The significance of absent end-diastolic velocity in umbilical artery velocity waveforms. Am J Obstet Gynecol 1987; 156: 1213-1218. [23] Fleischer A, Schulman H, Farmakides G, Bracero L, Blattner P, Randolph G. Umbilical artery velocity waveforms and intrauterine growth retardation. Am J Obstet Gynecol 1985; 151: 502-5. [24] Brar HS, Medearis AL, DeVare GR, Platt LD. Maternal and fetal blood flow velocity waveforms in patients with preterm labor: prediction of successful tocolysis. Am J Obstet Gynecol 1988; 159: 947-50. [25] Bilardo CM, Nicolaldes KH, Campbell S. Doppler measurements of fetal and uteroplacental circulations: relationship with umbilical venous blood gases measured at cordocentesis. Am J Obstet Gynecol 1987; 69: 705-9. [26] Wladimiroff JW, Wijngaard AGW, Degani S, Noordam M J, van Eyck J, Tonge HM. Cerebral and umbilical arterial blood flow velocity waveforms in normal and growth-retarded pregnancies. Obstet Oynecol 1987; 69: 705-9. [271 Beksa¢ MS, Onderoolu LS, Ozdemir K, Karaka~ O. The validation of a computerized system for the interpretation of the antepartum fetal heart rate tracings (version 89/2.34). Eur J Obstet Gynecol Reprod Biol 1991; 42: 9-14.