Learning Pelvic Anatomy: A Deep Learning Approach to Deformable Registration for Gynecologic Radiotherapy
Abstract
Purpose
To develop and validate a female pelvis-specific deep learning deformable image registration (DIR) framework optimized for longitudinal CT imaging, enabling accurate dose accumulation and cumulative dose assessment in gynecological radiotherapy.
Methods
Patients treated with pelvic radiotherapy between 2018 and 2024 were retrospectively identified. Patients with at least two CT scans were included, resulting in a dataset of 655 CT scans from 276 patients for model training, validation, and testing. Segmentations from TotalSegmentator and Ethos therapy (Varian), as well as supplemental data from the Amos dataset, were used create a custom nnUNet segmentation model for critical organs including uterus and rectum. Intra-patient affine registration was first performed using ANTs. Images were subsequently resampled to 1.5 mm3 isotropic spacing, cropped to the T12 vertebrae and lower body trunk, and zero padded to a uniform matrix size of 352 x 256 x 416 voxels. DIR was performed using a modified Voxelmorph framework. The model was adapted specifically for female pelvic anatomy, incorporating organ-aware supervision through auxiliary segmentation masks. A semi-supervised model was trained for intra-patient registration using paired CT scans.
Results
Our optimized registration model achieved a mean DICE score of 0.74, a mean Distance to Agreement of 3.25 mm, a Jacobian determinant of 0.002, and was driven by uterus, rectum, and urinary bladder segmentations. Best performance was obtained using a weighted loss combination of Normalized Cross Correlation (weight = 1.0), DICE Similarity Coefficient (weight = 0.2), and L2 regularization (weight = 0.1). The model was trained for [xxx] epochs at a learning rate of [xxx] using the Adam optimizer.
Conclusion
This study presents the first deep learning-based DIR framework specifically optimized for female pelvic CT imaging. By tailoring model architecture, loss functions and training strategy to female anatomy, this approach enables robust longitudinal registration across treatment plans.