Cross-Modal Transfer Learning Enables Clinically Useful Segmentation of Pediatric Brain Tumors using Diffusion Weighted Imaging

Motivation: Diffusion weighted imaging (DWI) maps have demonstrated excellent diagnostic utility in pediatric brain tumours through quantitative image analysis, which requires annotation of tumour boundaries. Automated segmentation models are not typically trained or deployed on DWI due to insufficient data availability. Goal(s): Determine whether cross-modal transfer learning (CMTL) can leverage large (non-DWI) public datasets to improve tumour segmentation performed on DWI alone. Approach: Trained deep-learning segmentation models with and without CMTL and assessed segmentation accuracy and their utility across multiple diagnostic tasks in pediatric brain tumours requiring DWI. Results: CMTL improves segmentation performance over baseline models, and maintains comparable diagnostic utility to manual annotations. Impact: Establishes benefits of leveraging large non-DWI public datasets, to improve automated DWI segmentation models, essential for native pediatric brain tumour analysis. This eliminates error arising from image co-registration, streamlines clinical workflows and limits the impact of missing imaging modalities.