ESTRO 2025 - Abstract Book

S2549

Physics - Autosegmentation

ESTRO 2025

Deep-learning based auto-segmentation of the GTV using mpMRI has the potential to improve consistency and reduce workload. However, generalizability of deep-learning models to other institutions is often limited. The aim of this work was to perform cross-institutional training and validation of a published deep-learning method [1], to segment prostate cancer GTVs with Gleason score (GS) ≥3+4 by using T2w and ADC images. Material/Methods: Two retrospective datasets were available. Center 1 data consisted of 112 patients with co-registered T2w and ADC images based on DWI with b-values of 0-800 and GTV delineations. Center 2 data consisted of 157 patients with co registered T2w and ADC images based on DWI with b-values 150-800 and GTV delineations validated by whole mount section histopathology. Each dataset was split for training-validation-test 70-15-15%, stratified for scanner type and GS. Four model instances for each centers’ data using ADC only and ADC + T2w inputs were trained to produce GTV segmentations. Postprocessing of models’ output were done to remove small islets (<1 cc). Evaluation was done in terms of recall, Dice similarity coefficient (DSC) of detected lesions and 95% Hausdorff distance (HD95). A lesion was defined as detected if DSC > 0.1. Results: Test data average GTV volume of centers 1 and 2 was 0.6 (IQR: 0.5-0.7) cc and 1.4 (IQR: 0.9-2.8) cc respectively. On evaluation, average prediction volume was 1.5 (IQR: 0.7-2.0) cc and 1.8 (IQR: 0.8-2.6) cc respectively. For both test sets, recall was highest for the ADC only model from center 2 (0.72-0.83, figure 1). Focusing on detected lesions only, DSC and HD95 showed consistent results over all models and datasets (median DSC range 0.16-0.36; median HD95 range 4.01-6.00 mm), with no single model clearly outperforming the rest. Improved performance on center 2 data may be due to larger GTV volumes.