Clinically aligned whole-body MRI segmentation of skeletal metastases via Supervised Anatomical Pretraining.

In oncology practice, response assessment of metastatic disease requires reliable and reproducible quantification of measurable metastatic burden. Manual identification, segmentation, and volumetry of all lesions is labor-intensive and variable, limiting routine clinical adoption. An automated approach is therefore needed. Segmenting metastatic bone disease (MBD) on whole-body MRI (WB-MRI) is challenging because of the heterogeneous appearance and anatomical distribution of lesions, ambiguous boundaries, and the low volumetric prevalence of metastatic deposits within the body. Training robust machine learning models for this task requires large, well-annotated datasets that capture lesion variability. However, assembling such datasets demands substantial expert time and is prone to annotation error. Although self-supervised learning (SSL) can take advantage of large unlabeled datasets, the learned representations tend to remain generic and may miss the subtle anatomical and pathological features essential for accurate lesion detection. In this work, we propose a Supervised Anatomical Pretraining (SAP) method that learns from a limited dataset of anatomical labels. First, an MRI-based skeletal segmentation model is developed and trained on WB-MRI scans from healthy individuals for high-quality skeletal delineation. Then, we compare its downstream efficacy in segmenting MBD on a cohort of 40 patients with metastatic prostate cancer, against a randomly initialized baseline and a state-of-the-art self-supervised method. SAP significantly outperforms both the Baseline and SSL-pretrained models achieving a normalized surface Dice of 0.78 and a Dice coefficient of 0.66. The method achieved a lesion detection