AI-Based Auto-Segmentation for Loco-Regional Breast Cancer Radiotherapy: Training, Validation and Clinical Implementation

Abstract

Listen

Translate article

<h3>Purpose/Objective(s)</h3> To train and validate a comprehensive deep-learning (DL) segmentation model for loco-regional breast cancer with the aim of clinical implementation. <h3>Materials/Methods</h3> 200 left-sided breast cancer cases from two radiotherapy centers in Norway were included in the project. In all cases, 7 clinical target volumes (CTVs) were manually delineated based on the ESTRO guidelines, in addition to 11 organs at risk (OARs). 170 patient cases were used for training, which involved neural networks of 3D CNN U-net type. The remaining 30 cases were used for model validation, which included: 1) evaluation of geometric similarity between DL structures and manually delineated structures using 95-percentile Hausdorff distance (HD95) and Dice similarity coefficient (DSC), benchmarked against inter-observer variation (IOV), 2) qualitative scoring using a 4-step scale according to clinical usability, and 3) dosimetric analysis of conformal treatment plans which were based on uncorrected DL structures. A dedicated team of 3 experienced oncologists (EO) and 3 experienced radiation therapists contributed to the manual delineations for model training and validation, while 5 EOs delineated the same 5 patient cases for evaluation of IOV. To test for difference in model performance and IOV, Mann-Whitney U test was used, while paired sample T-test was used for the dosimetric analysis of OARs. <h3>Results</h3> Geometrically, the model performed better than IOV for all CTVs and most OARs, see Table 1 for main results. Qualitatively, no or only minor corrections were required for 14 % and 71 % of the CTVs and 72 % and 26 % of the OARs, respectively. Major corrections were required for 15 % of the CTVs and 2 % of the OARs, while no structure scored as "not usable". The most frequent corrections were adjustments in the cranial and caudal parts of the structures. Dosimetrically, the CTV coverage, based on D98 > 95 %, was fulfilled for 100 % and 89 % of the breast and lymph node structures, respectively. No differences in OAR dose parameters were considered clinically relevant. The model was implemented in a commercial treatment planning system, which generates the structures in 1.5 minutes. <h3>Conclusion</h3> Convincing results from the validation process led to the decision of clinical implementation. The model is now in clinical use, and will be monitored regarding applicability, standardization and efficiency. Table 1. Geometric performance of the DL model and corresponding IOV for all CTVs and two of the OARs (median values). Bold type indicates the best metric, and underscore indicate statistical significance (p<0.05).