Deep ensemble optimized models for probabilistic CTV breast segmentation

To optimize radiotherapy treatment and minimize toxicities, effective segmentation of organs-at-risk (OARs) and clinical target volume (CTV) is essential. Deep learning (DL) models can achieve high segmentation accuracy through careful tuning. However, their reliability also hinges on addressing uncertainties stemming from variability in clinical contouring practices. This study systematically evaluates six advanced DL models for automatic CTV segmentation in whole-breast radiotherapy. It further leverages the top-performing models to construct a probability map, aiming to improve consistency and mitigate bias in clinical predictions.

Planning CTs from a single institute (861: training, 100: "temporal" test) were used to train six models for right/left breast CTV segmentation simultaneously: UNet, SegResNetDS, DynUNet (MONAI), nnU-Net (Total Segmentator), MedSAM2 (MS-A: CT-specific weights, MS-B: general medical-image weights). MedSAM2 employed a strong caveat: it needs a spatial prompt to make the prediction. Standard metric such as Dice Similarity Coefficient, Average Surface Distance and Hausdorff Distance were used to evaluate prediction compared to ground truth. A Friedman test followed by post-hoc pairwise comparisons through Conover test were conducted on the temporal test set. In addition, the best models derived (in line with inter-observer variability IOV) were used to build model-based probability maps and quantify the differences between high concordance (100%) and lower concordance (25%) isoprobabilities to clinician CTV.

All models attained overall satisfactory performance [comparable to IOV (Dice = 0.90)]. Among them, UNet, DynUNet, nnU-Net, and MedSAM2-A demonstrated the highest equivalent accuracy, with average ASD = 1.5 mm and HD95 = 3.8 mm. Conover test displayed more subtle differences with lower performance of two models out of six (SegResNetDS, MS-B), which were therefore removed for the subsequent built of a probability map. The analysis of the 100 vs. 25% isoprobability volumes on the temporal test dataset (average difference = 123 ± 81 cm³) highlighted areas of greater uncertainty at the lateral and cranio-caudal CTV borders.

Different DL architectures were optimized, trained and validated on a large cohort of patients, successfully predicting the CTV for both right/left breast. In addition, the first attempt to provide uncertainty maps was achieved through successful computation of DL model-based probability maps. Supported by CCM 2024 (Ministry of Health).