Enhancing Cross-Modal Medical Image Segmentation through Compositionality
This addresses segmentation problems for medical imaging applications where different modalities produce varying image characteristics, representing an incremental improvement with novel method elements.
The paper tackles the challenge of cross-modal medical image segmentation by introducing compositionality as an inductive bias in a segmentation network, using learnable von Mises-Fisher kernels for content-style disentanglement. The results show enhanced segmentation performance and reduced computational costs on multiple medical datasets.
Cross-modal medical image segmentation presents a significant challenge, as different imaging modalities produce images with varying resolutions, contrasts, and appearances of anatomical structures. We introduce compositionality as an inductive bias in a cross-modal segmentation network to improve segmentation performance and interpretability while reducing complexity. The proposed network is an end-to-end cross-modal segmentation framework that enforces compositionality on the learned representations using learnable von Mises-Fisher kernels. These kernels facilitate content-style disentanglement in the learned representations, resulting in compositional content representations that are inherently interpretable and effectively disentangle different anatomical structures. The experimental results demonstrate enhanced segmentation performance and reduced computational costs on multiple medical datasets. Additionally, we demonstrate the interpretability of the learned compositional features. Code and checkpoints will be publicly available at: https://github.com/Trustworthy-AI-UU-NKI/Cross-Modal-Segmentation.