- Email: [email protected]
EP-1855: Computed Tomography lung texture changes due to radiotherapy for non-small cell lung cancer
ESTRO 35 2016
S873
________________________________________________________________________________
Conclusion: This study indicates that both glucose metabolism measured by PET/CT and restriction measured by DW-MR are independent cellular phenomena in newly diagnosed esophageal cancer. Therefore, SUV with ADC values may have complementary roles as imaging biomarkers in the evaluation of survival and response to neoadjuvant treatment in esophageal cancer. EP-1854 Mammographic texture features for determination breast cancer molecular subtype M. Arenas Prat1, L. Díez-Presa1, J. Torrents-Barrena2, M. Arquez1, C. Pallas1, M. Gascón1, M. Bonet1, A. Latorre-Musoll3, S. Sabater4, D. Puig2 1 Hospital Universitari Sant Joan de Reus, Radiation Oncology, Reus, Spain 2 University Rovira i Virgili, Computer Science and Mathematics, Tarragona, Spain 3 Hospital de la Santa Creu i Sant Pau, Physics, Barcelona, Spain 4 Complejo Hospitalario Universitario, Radiation Oncology, Albacete, Spain Purpose or Objective: To determine molecular subtypes of breast cancers using texture-feature-driven machine learning techniques on mammographic images. Material and Methods: We used mammograms of 61 ductal carcinomas (grade 2-3, median age 60, mean tumor size 28mm). A physician defined a 100x100 ROI around tumors on mammographic images. Extraction of texture features was performed using three independent descriptors: Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG) and Gabor Filter (GF). Then, a supervised classification was applied using two independent classifiers: K-Nearest Neighbors (KNN) and Support Vector Machines (SVM) (both linear- and radial-type). Both classifiers were trained to identify the molecular subtype (Luminal A, Luminal B (Her2-), Luminal B (Her2+), Her2+, Basal Like and carcinoma in situ) using the first 38 mammograms. We assessed the accuracy of our machine learning technique using the last 23 mammograms.
Results: Accuracy of SVM-R classifier was 52% irrespective of the texture descriptor we used. SVM-L/KNN classifiers achieved an accuracy of 48/39, 35/30 and 21/35% for LBP, HOG and GF descriptors. When simplifying the classification problem to only two subtypes, Luminal A and Luminal B (Her2-), classifier accuracies astonishingly improved. SVM-R accuracy was 75% irrespective of the texture descriptor and SVM(L)/KNN accuracies were 38/75, 50/50 and 63/75% for LBP, HOG and GF. Conclusion: Our texture-feature-driven machine learning technique provides a reliable classification into molecular subtypes using mammographic images only. Accuracy improves when simplifying to only two subtypes. We expect even better accuracies by increasing the number of patients used for the training stage of our machine learning technique. EP-1855 Computed Tomography lung texture changes due to radiotherapy for non-small cell lung cancer J. Chalubinska-Fendler1, W. Fendler2, Ł. Karolczak3, C. Chudobiński4, J. Łuniewska-Bury5, A. Materka3, J. Fijuth1 1 Medical University of Lodz, Radiotherapy Department- Chair of Oncology, Lodz, Poland 2 Medical University of Lodz, Department of PediatricsOncology- Hematology and Diabetology, Lodz, Poland 3 Lodz University of Technology, Institute of Electronics, Lodz, Poland 4 Regional Oncological Centre- Lodz, Radiology Department, Lodz, Poland 5 Regional Oncological CentreLodz, Brachytherapy Department, Lodz, Poland Purpose or Objective: Radiation induced lung toxicity (RILT) may occur in 5-20% of patients irradiated due to Non-Small Cell Lung Cancer (NSCLC) but may be asymptomatic during the course of radiotherapy (RTx). Computed tomography (CT) image changes induced by RILT present after 3-9 months since RTx, mostly as lung fibrosis. Early changes on lung tissue image, i.e. during treatment, are not possible to diagnose by the naked eye, but could be detect by computerassisted texture analysis. Material and Methods: Fifteen patients aged 63.7+/-6.4 years, with NSCLC undergoing RTx were tested using CT before RTx and after receiving 40Gy of dose prescribed to PTV. Images were entered into a texture analysis program – MaZda® which extracted 284 texture parameters based on: signal intensity, variability of signal intensity, autocorrelation, direction of change, Fourier spectrum, Wavelet spectrum and repeatability of intensity change patterns. From every patient 10 regions of interest (ROIs)
S874 ESTRO 35 2016 _____________________________________________________________________________________________________ were extracted (see image) at both timepoints from two sections of lung tissue – one that received the highest planned dose in healthy tissue and one that received low or no dose of RTx. Linear discriminant analysis (LDA) with 5-fold cross-validation and backward stepwise selection of variables was used to construct best classification models to separate irradiated from non-irradiated regions of the lung and differentiation of patients with RILT and without.
Results: LDA based on seven parameters allowed for differentiation (area under the ROC curve 0.86) of regions of healthy tissue and regions from tumour’s site. The 7 variables were Wavelet-transform functions of different frequencies. However, despite those differences, CT-images from the 40Gy timepoints differed significantly depending on the received dose. An LDA model based on six parameters (3 autocorrelation functions, kurtosis and 2 wavelet function parameters) differentiated non-irradiated regions from irradiated ones – ROC AUC 0.89 (95%CI 0.75-1.00). Preliminary data from follow up showed that patients in whom RILT developed (N= 7) could be differentiated from those free from complications (ROC 0.96 95%CI 0.89-1.00). However, parameters used in this LDA-based classifier relied on CT texture parameters extracted from both irradiated and non-irradiated ROIs, making ROI selection a crucial part of the texture analysis process.
Material and Methods: PET delineated tumors for 158 NSCLC patients were used to calculate textural features as proposed by Amadasun et al [IEEE Trans. Syst., Man, Cybern., Syst. 1989]. Delineations of the tumors were made semiautomatically based on EANM guidelines for VOI41 and VOI50 (delineation at the 41% and 50% level of SUVmax). Additional smoothened delineations were made to resemble delineations made by humans. Furthermore, dilated versions of VOI41 were analyzed to determine the response of the textural features to large changes in delineation. Textural features are typically used to divide a patient population in two groups based on a given textural cut-value, e.g. the median value. Thus, the textural feature should preferably be stable towards small delineation variations in terms of patient classification. Such stability was tested using ROC curves, in which the initial delineation (VOI41) was used as ground truth classification based on the median value. Volume dependence of the textural features was assessed through the Spearman correlation coefficient. Results: Coarseness, busyness, contrast, and complexity were all confirmed to have a significant correlation with volume (absolute Spearman > 0.58). The figure shows coarseness’ ability to classify the patients consistently for different delineations. The large area under the ROC curve (almost unity) between VOI41, VOI50, and the smoothed VOI, shows that the patient classification is almost independent of small variations of delineation. The figure also shows how successive dilations of tumor volume reduce the area under the curve. Similar findings were observed for the textures busyness and contrast. A mathematical examination of the textural features provided an easy way to correct for the volume dependence of coarseness and contrast. Neither of these modified versions was found to have volume dependence (absolute Spearman < 0.22); while at the same time having the same stability characteristic as their original versions. Conclusion: All original textures had strong correlation with volume, which for PET delineation of lung tumors could be a confounding factor within a textural predictor. Through small changes to the original definition it is possible to make coarseness and contrast volume independent; a property which is needed for the features to be used as independent predictive factors.
Conclusion: Texture of CT-scans contains enough information to detect RTx-induced changes, although the method may be affected by pretreatment differences, which necessitates a robust placement of ROIs for analysis. EP-1856 Predictive factors based on textural features – reliability of patient classification T.L. Jacobsen1, U. Bernchou2, T. Schytte3, O. Hansen3, C. Brink2 1 University of Southern Denmark, Department of PhysicsChemistry- and Pharmacy, Odense, Denmark 2 Odense University Hospital, Laboratory of Radiation Physics, Odense, Denmark 3 Odense University Hospital, Department of Oncology, Odense, Denmark Purpose or Objective: Textural analysis of lung tumors in PET or CT images is currently of interest in a number of publications e.g. to predict overall survival after radiotherapy. Given that tumor volume is a known independent predictor in radiotherapy of lung cancer, a textural feature must be volume independent to gain independent predictive power. Furthermore, the feature value should be stable against small variations in the delineated tumor volume. This study analyses how changes in PET based tumor volume and delineation affect different published textural features.
EP-1857 Multi-parametric MRI at 3.0 Tesla for the prediction of treatment response in rectal cancer T. Pham1,2,3, G. Liney1,4,5, K. Wong1,2,3, D. Roach6, D. Moses2,7, C. Henderson2,8,9, M. Lee1, R. Rai1, M. Barton1,2,3
S873
________________________________________________________________________________
Conclusion: This study indicates that both glucose metabolism measured by PET/CT and restriction measured by DW-MR are independent cellular phenomena in newly diagnosed esophageal cancer. Therefore, SUV with ADC values may have complementary roles as imaging biomarkers in the evaluation of survival and response to neoadjuvant treatment in esophageal cancer. EP-1854 Mammographic texture features for determination breast cancer molecular subtype M. Arenas Prat1, L. Díez-Presa1, J. Torrents-Barrena2, M. Arquez1, C. Pallas1, M. Gascón1, M. Bonet1, A. Latorre-Musoll3, S. Sabater4, D. Puig2 1 Hospital Universitari Sant Joan de Reus, Radiation Oncology, Reus, Spain 2 University Rovira i Virgili, Computer Science and Mathematics, Tarragona, Spain 3 Hospital de la Santa Creu i Sant Pau, Physics, Barcelona, Spain 4 Complejo Hospitalario Universitario, Radiation Oncology, Albacete, Spain Purpose or Objective: To determine molecular subtypes of breast cancers using texture-feature-driven machine learning techniques on mammographic images. Material and Methods: We used mammograms of 61 ductal carcinomas (grade 2-3, median age 60, mean tumor size 28mm). A physician defined a 100x100 ROI around tumors on mammographic images. Extraction of texture features was performed using three independent descriptors: Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG) and Gabor Filter (GF). Then, a supervised classification was applied using two independent classifiers: K-Nearest Neighbors (KNN) and Support Vector Machines (SVM) (both linear- and radial-type). Both classifiers were trained to identify the molecular subtype (Luminal A, Luminal B (Her2-), Luminal B (Her2+), Her2+, Basal Like and carcinoma in situ) using the first 38 mammograms. We assessed the accuracy of our machine learning technique using the last 23 mammograms.
Results: Accuracy of SVM-R classifier was 52% irrespective of the texture descriptor we used. SVM-L/KNN classifiers achieved an accuracy of 48/39, 35/30 and 21/35% for LBP, HOG and GF descriptors. When simplifying the classification problem to only two subtypes, Luminal A and Luminal B (Her2-), classifier accuracies astonishingly improved. SVM-R accuracy was 75% irrespective of the texture descriptor and SVM(L)/KNN accuracies were 38/75, 50/50 and 63/75% for LBP, HOG and GF. Conclusion: Our texture-feature-driven machine learning technique provides a reliable classification into molecular subtypes using mammographic images only. Accuracy improves when simplifying to only two subtypes. We expect even better accuracies by increasing the number of patients used for the training stage of our machine learning technique. EP-1855 Computed Tomography lung texture changes due to radiotherapy for non-small cell lung cancer J. Chalubinska-Fendler1, W. Fendler2, Ł. Karolczak3, C. Chudobiński4, J. Łuniewska-Bury5, A. Materka3, J. Fijuth1 1 Medical University of Lodz, Radiotherapy Department- Chair of Oncology, Lodz, Poland 2 Medical University of Lodz, Department of PediatricsOncology- Hematology and Diabetology, Lodz, Poland 3 Lodz University of Technology, Institute of Electronics, Lodz, Poland 4 Regional Oncological Centre- Lodz, Radiology Department, Lodz, Poland 5 Regional Oncological CentreLodz, Brachytherapy Department, Lodz, Poland Purpose or Objective: Radiation induced lung toxicity (RILT) may occur in 5-20% of patients irradiated due to Non-Small Cell Lung Cancer (NSCLC) but may be asymptomatic during the course of radiotherapy (RTx). Computed tomography (CT) image changes induced by RILT present after 3-9 months since RTx, mostly as lung fibrosis. Early changes on lung tissue image, i.e. during treatment, are not possible to diagnose by the naked eye, but could be detect by computerassisted texture analysis. Material and Methods: Fifteen patients aged 63.7+/-6.4 years, with NSCLC undergoing RTx were tested using CT before RTx and after receiving 40Gy of dose prescribed to PTV. Images were entered into a texture analysis program – MaZda® which extracted 284 texture parameters based on: signal intensity, variability of signal intensity, autocorrelation, direction of change, Fourier spectrum, Wavelet spectrum and repeatability of intensity change patterns. From every patient 10 regions of interest (ROIs)
S874 ESTRO 35 2016 _____________________________________________________________________________________________________ were extracted (see image) at both timepoints from two sections of lung tissue – one that received the highest planned dose in healthy tissue and one that received low or no dose of RTx. Linear discriminant analysis (LDA) with 5-fold cross-validation and backward stepwise selection of variables was used to construct best classification models to separate irradiated from non-irradiated regions of the lung and differentiation of patients with RILT and without.
Results: LDA based on seven parameters allowed for differentiation (area under the ROC curve 0.86) of regions of healthy tissue and regions from tumour’s site. The 7 variables were Wavelet-transform functions of different frequencies. However, despite those differences, CT-images from the 40Gy timepoints differed significantly depending on the received dose. An LDA model based on six parameters (3 autocorrelation functions, kurtosis and 2 wavelet function parameters) differentiated non-irradiated regions from irradiated ones – ROC AUC 0.89 (95%CI 0.75-1.00). Preliminary data from follow up showed that patients in whom RILT developed (N= 7) could be differentiated from those free from complications (ROC 0.96 95%CI 0.89-1.00). However, parameters used in this LDA-based classifier relied on CT texture parameters extracted from both irradiated and non-irradiated ROIs, making ROI selection a crucial part of the texture analysis process.
Material and Methods: PET delineated tumors for 158 NSCLC patients were used to calculate textural features as proposed by Amadasun et al [IEEE Trans. Syst., Man, Cybern., Syst. 1989]. Delineations of the tumors were made semiautomatically based on EANM guidelines for VOI41 and VOI50 (delineation at the 41% and 50% level of SUVmax). Additional smoothened delineations were made to resemble delineations made by humans. Furthermore, dilated versions of VOI41 were analyzed to determine the response of the textural features to large changes in delineation. Textural features are typically used to divide a patient population in two groups based on a given textural cut-value, e.g. the median value. Thus, the textural feature should preferably be stable towards small delineation variations in terms of patient classification. Such stability was tested using ROC curves, in which the initial delineation (VOI41) was used as ground truth classification based on the median value. Volume dependence of the textural features was assessed through the Spearman correlation coefficient. Results: Coarseness, busyness, contrast, and complexity were all confirmed to have a significant correlation with volume (absolute Spearman > 0.58). The figure shows coarseness’ ability to classify the patients consistently for different delineations. The large area under the ROC curve (almost unity) between VOI41, VOI50, and the smoothed VOI, shows that the patient classification is almost independent of small variations of delineation. The figure also shows how successive dilations of tumor volume reduce the area under the curve. Similar findings were observed for the textures busyness and contrast. A mathematical examination of the textural features provided an easy way to correct for the volume dependence of coarseness and contrast. Neither of these modified versions was found to have volume dependence (absolute Spearman < 0.22); while at the same time having the same stability characteristic as their original versions. Conclusion: All original textures had strong correlation with volume, which for PET delineation of lung tumors could be a confounding factor within a textural predictor. Through small changes to the original definition it is possible to make coarseness and contrast volume independent; a property which is needed for the features to be used as independent predictive factors.
Conclusion: Texture of CT-scans contains enough information to detect RTx-induced changes, although the method may be affected by pretreatment differences, which necessitates a robust placement of ROIs for analysis. EP-1856 Predictive factors based on textural features – reliability of patient classification T.L. Jacobsen1, U. Bernchou2, T. Schytte3, O. Hansen3, C. Brink2 1 University of Southern Denmark, Department of PhysicsChemistry- and Pharmacy, Odense, Denmark 2 Odense University Hospital, Laboratory of Radiation Physics, Odense, Denmark 3 Odense University Hospital, Department of Oncology, Odense, Denmark Purpose or Objective: Textural analysis of lung tumors in PET or CT images is currently of interest in a number of publications e.g. to predict overall survival after radiotherapy. Given that tumor volume is a known independent predictor in radiotherapy of lung cancer, a textural feature must be volume independent to gain independent predictive power. Furthermore, the feature value should be stable against small variations in the delineated tumor volume. This study analyses how changes in PET based tumor volume and delineation affect different published textural features.
EP-1857 Multi-parametric MRI at 3.0 Tesla for the prediction of treatment response in rectal cancer T. Pham1,2,3, G. Liney1,4,5, K. Wong1,2,3, D. Roach6, D. Moses2,7, C. Henderson2,8,9, M. Lee1, R. Rai1, M. Barton1,2,3