Can Deep Learning-Based Auto-Contouring Software Achieve Accurate Pelvic Volume Delineation in Volumetric Image-Guided Radiotherapy for Prostate Cancer? A Preliminary Multicentric Analysis.

Abstract

Background: Radiotherapy (RT) is a mainstay treatment for prostate cancer (PC). Accurate delineation of organs at risk (OARs) is crucial for optimizing the therapeutic window by minimizing side effects. Manual segmentation is time-consuming and prone to inter-operator variability. This study investigates the performance of Limbus^® Contour^® (LC), a deep learning-based auto-contouring software, in delineating pelvic structures in PC patients.

Methods: We evaluated LC's performance on key structures (bowel bag, bladder, rectum, sigmoid colon, and pelvic lymph nodes) in 52 patients. We compared auto-contoured structures with those manually delineated by radiation oncologists using different metrics.

Results: LC achieved good agreement for the bladder (median Dice: 0.95) and rectum (median Dice: 0.83). However, limitations were observed for the bowel bag (median Dice: 0.64) and sigmoid colon (median Dice: 0.6), with inclusion of irrelevant structures. While the median Dice for pelvic lymph nodes was acceptable (0.73), the software lacked sub-regional differentiation, limiting its applicability in certain other oncologic settings.

Conclusions: LC shows promise for automating OAR delineation in prostate radiotherapy, particularly for the bladder and rectum. Improvements are needed for bowel bag, sigmoid colon, and lymph node sub-regionalization. Further validation with a broader and larger patient cohort is recommended to assess generalizability.