ESTRO 2020 Abstract book

S896 ESTRO 2020

PO-1562 Radiomics applied to dose distributions to predict toxicity after radiotherapy in cervical cancer F. Lucia 1 , V. Bourbonne 1 , G. Dissaux 1 , O. Miranda 1 , G. Dissaux 1 , U. Schick 1 , D. Visvikis 2 , J. Bert 2 , M. Hatt 2 , R. Abgral 3 1 Radiation Oncology department- University Hospital- Brest- France, Radiation Oncology department- University Hospital- Brest- France, Brest, France ; 2 LaTIM- INSERM- UMR 1101- Univ Brest- Brest- France, LaTIM- INSERM- UMR 1101- Univ Brest- Brest- France, Brest, France ; 3 Nuclear Medicine department- University Hospital- Brest- France, Nuclear Medicine department- University Hospital- Brest- France, Brest, France Purpose or Objective Dose-volume histograms (DVH) do not account for the spatial distributions of dose at the voxel level. Our goal was to compare the prediction of toxicity events after radiotherapy for locally advanced cervical cancer (LACC), relying on either DVH parameters or the use of the radiomics approach applied todose maps at the voxel level. Material and Methods Acute and late toxicity scores using the CTCAE v4, spatial dose distributions, and usual clinical predictors for toxicity, such as age and baseline symptoms, of 102 patients treated with chemoradiotherapy followed by brachytherapy for LACC were used in this study. Patients were split into training (patients treated in Brest, n=52) and testing (patients treated in Quimper, n=50) sets. In addition to usual DVH parameters, 91radiomic features (intensity and texture) of rectum, bladder and vaginal 3D dose distributions (defined using the organs at risk contours from the planning) were extracted after discretization into fixed bin width of 55 Gy, and evaluated for predictive modelling of gastrointestinal (GI), genitourinary (GU) and vaginal toxicities. Logistic Normal Tissue Complication Probability (NTCP) models were derived, using only clinical parameters, clinical + DVH, clinical + radiomics, DVH + radiomics and clinical + DVH + radiomics combinations. Results In the entire cohort (training/testing), the cumulative rates of acute and late grade ≥2 GU, GI and vaginal toxicities were 39%/42%, 14%/16%, 26%/28%, 10%/14%, 24%/24%, and 24%/24%, respectively. In the training cohort, for GI acute/late toxicities, the area under the curve (AUC) using only clinical parameters was 0.53/0.65, which increased to 0.66/0.63, and 0.76/0.87, when adding DVH or radiomics parameters respectively. For GU acute/late toxicities, the AUC went up from 0.55/0.56 (clinical only), to 0.84/0.90 (+DVH) and 0.83/0.96 (clinical + DVH + radiomics). For vaginal acute/late toxicities, the AUC went up from 0.51/0.57 (clinical only), to 0.58/0.57 (+DVH) and 0.82/0.79 (clinical + DVH + radiomics). These trends were also confirmed by the AUCs of the final models evaluated in the testing cohort. For GI acute/late toxicities, the AUC for using only clinical parameters was 0.55/0.56, which increased to 0.56/0.55, and 0.78/0.85, when adding DVH or radiomics parameters respectively. For GU acute/late toxicities, the AUC went up from 0.55/0.67 (clinical only), to 0.62/0.67 (+DVH) and 0.72/0.73 (clinical + DVH + radiomics). For vaginal acute/late toxicities, the AUC went up from 0.54/0.58 (clinical only), to 0.53/0.53 (+DVH) and 0.74/0.76 (clinical + DVH + radiomics). Conclusion The predictive performance of NTCP models based on radiomics features is better than the commonly used DVH parameters. Dosimetric radiomics analysis is a promising tool for NTCP modelling in radiotherapy.

co-occurrence (24), gray level run length (16), gray level size zone (16), neighboring gray tone difference (5), and gray level dependence (14) matrices. Eight wavelet decompositions were obtained using a coiflet wavelet transformation; each decomposition was used as an input image to calculate the first-order statistics and the textural features described above. To tackle the problem of unbalanced training set, Synthetic Minority Over-sampling Technique (SMOTE) was used to create synthetic data making the two categories with equal number of samples. Given the high number of features in comparison to the number of patients, feature preselection was performed to reduce oversampling. Three algorithms were tested: the minimum redundancy maximum relevance (MRMR), mutual information score (MI) and Anova F-value score (AF).42 features were preselected. The classification approach was Random Forest (RF).RF’s parameters were optimized by cross- validated grid-search over a parameter grid. Three machine learning models (MLM) were built: (MRMR, RF), (MI, RF), and (AF, RF). The performance of MLM was assessed by the computation of the ROC curve, the area under the ROC curve (AUC- measure of separability between classes for any threshold), the accuracy (ACC- the fraction of correct predictions), the precision (PR- proportion of positive identifications that was actually correct) and the recall (RE- proportion of actual positives that was identified correctly). Results Table 1 displays AUC, ACC, PR and RE, on the validation set, of optimized RF for each feature selection approach. AUC, ACC, PR, RE were close to 1 for the training set. Fig 1 shows the roc curves for each combination (MRMR, RF), (AF, RF) and (MI, RF). The best performance of RF was obtained with MRMR feature selection.

Conclusion We demonstrated that RF trained on preselected radiomic features could potentially predict the stage of NSCLC. Further investigation with different MLM and more data will be done.