A lightweight deep learning model for automated segmentation of Gynecologic organs and cervical Tumors on T2-weighted magnetic resonance imaging

Why clinicians should know about it

• Picked for Medical Physics (paper of the day, 11 September 2026): Lightweight DL model for automated cervical MRI segmentation

Abstract

Background and Purpose: Manual segmentation of gynecologic organs and cervical tumors for radiotherapy planning is time-consuming and variable. Automated segmentation on routine T2-weighted magnetic resonance imaging (MRI) remains limited. The aim of this study was to evaluate a lightweight deep learning model for automated segmentation of gynecologic organs and cervical tumors on T2-weighted MRI. Materials and Methods: This work applied a two-stage lightweight deep learning model (PocketNet) to segment the cervix, vagina, uterus, and tumor(s) on T2-weighted MRI in 102 patients with cervical cancer undergoing definitive radiotherapy. Model performance was assessed using the Dice-Sorensen coefficient (DSC) and 95th percentile Hausdorff distance (Haus95) on internal data and validated on an external dataset. A full nnU-Net model trained on the same internal dataset served as a benchmark for segmentation accuracy and computational efficiency. Results: On the institutional dataset, PocketNet achieved mean DSC values exceeding 70% for tumor segmentation and 80% for organ segmentation. External validation on The Cancer Imaging Archive (TCIA) Cervical Cancer Tumor Heterogeneity (CCTH) collection demonstrated the model's robustness, achieving DSC scores of 67.3% for tumor segmentation and 80.8% for organ segmentation. Compared to the PocketNet architecture, nnUNet achieved similar accuracy but required approximately twice the training time, more than 35 times as many parameters, and 40 times more memory for model storage. Conclusion: The PocketNet architecture provides reliable automated segmentation of gynecologic organs and cervical tumors on T2-weighted MRI, with performance comparable to a full-sized nnUNet while requiring substantially less memory and training time, supporting its potential integration into time-sensitive radiotherapy workflows.

Abstract as published, via PubMed.