Hybrid deep learning enables multi-institutional delineation of active bone marrow for gynecologic radiotherapy.

Background And Purpose: Pelvic radiotherapy for gynecologic cancer inevitably irradiates sensitive areas like iliac bones, lumbar vertebrae, and sacrum. Using F-FDG PET/CT as a reference, we developed a deep learning method to detect hematopoietic active bone marrow (ABM) on CT in gynecologic cancer patients and assess clinical benefits.

Materials And Methods: We analyzed 319 patients from five institutions retrospectively. ABM was divided into three 18F-FDG PET/CT-defined subregions: active iliac bone marrow (A_IBM), active sacral bone marrow (A_SBM), and active lumbar vertebrae bone marrow (A_LVBM), defined as areas with standardized uptake values exceeding subregional means. Six deep learning models were trained: hybrid nnU-Net, U-Net, V-Net, ResU-Net, nnU-Net, and UNETR. The hybrid nnU-Net approach integrated nnU-Net predictions with anatomical bone structures via Boolean operations, providing a post-processing strategy. The dataset was split into 290 cases for training and 29 for independent testing. Performance was evaluated using Dice similarity coefficients (DSCs) and 95th percentile Hausdorff distance (HD95). Two clinical cases were prospectively evaluated for ABM-sparing radiotherapy with hematologic monitoring.

Results: The hybrid nnU-Net achieved the highest DSCs for A_IBM (0.74 ± 0.06), A_LVBM (0.79 ± 0.07), and A_SBM (0.75 ± 0.06), with significant improvements over most models (p < 0.001), except nnU-Net. Despite ResU-Net's lower HD95 in two subregions, hybrid nnU-Net showed superior accuracy. No grade ≥2 hematologic toxicity occurred in prospective cases.

Conclusion: This multi-institutional study confirms that the hybrid nnU-Net accurately segments ABM from CT images, showing potential for ABM-sparing radiotherapy.