Modality-AGnostic image Cascade (MAGIC) for multi-modality cardiac substructure segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Cardiac substructure delineation is crucial for minimizing radiation-induced heart disease during treatment planning.
• Deep learning models offer potential for automating contouring but often lack generalizability across different imaging modalities and overlapping structures.

Purpose of the Study:

• To introduce and validate the Modality-AGnostic Image Cascade (MAGIC) deep-learning pipeline for comprehensive, multi-modal cardiac substructure segmentation.
• To assess MAGIC's performance in segmenting twenty cardiac substructures across various imaging modalities.

Main Methods:

• The MAGIC pipeline utilizes replicated encoding and decoding branches of an nnU-Net backbone to process multi-modality inputs and handle overlapping labels.
• The model was trained on a semi-supervised dataset (n=151) from the multi-modality whole-heart segmentation (MMWHS) dataset, including cardiac CT-angiography (CCTA) and MR modalities.
• Performance was evaluated using Dice Similarity Coefficient (DSC) and compared against fourteen single-modality baseline models.

Main Results:

• MAGIC achieved high average MMWHS DSC scores (0.88 ± 0.08 for CCTA, 0.87 ± 0.04 for MR), outperforming unimodal baselines.
• Average 20-structure DSC scores varied by modality, with CCTA yielding the highest overall performance (0.80 ± 0.16).
• The pipeline demonstrated significant reductions in training time (>80%) and parameters (>70%) compared to baseline models.

Conclusions:

• MAGIC provides an efficient and lightweight solution for segmenting cardiac substructures across multiple imaging modalities and overlapping structures within a single model.
• The pipeline achieves high segmentation accuracy without compromising performance, offering a valuable tool for radiotherapy planning.
• MAGIC's generalizability across modalities addresses a key limitation of current deep learning approaches in medical image segmentation.