SwitchNet: Adaptive Distribution Switching in UNet for Brain Lesion Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Automatic brain lesion segmentation aids diagnosis through texture analysis and tumor subregion delineation.
• Multimodal MRI improves segmentation by integrating complementary data, but conventional methods lack interpretability or adaptability.

Purpose of the Study:

• To introduce SwitchNet, a novel model for interpretable and parameter-efficient multimodal MRI brain lesion segmentation.
• To address limitations of existing methods regarding modality contribution interpretability and adaptability to varying modalities.

Main Methods:

• Proposed Adaptive Encoder and Decoder Blocks with dynamic switching for efficient feature allocation and modality utilization.
• Introduced a Guide-Contribution Mechanism for quantitative insights into individual modality contributions during segmentation training.

Main Results:

• SwitchNet achieved competitive segmentation performance on benchmark datasets (BraTS 2023, ISLES 2022, UCSF-PDGM).
• Demonstrated significantly enhanced interpretability and maintained parameter efficiency without additional computational cost.
• Validated the model's ability to automatically identify and leverage informative modalities for segmentation.

Conclusions:

• SwitchNet offers an interpretable and efficient solution for multimodal brain lesion segmentation.
• The model's dynamic switching and contribution mechanism provide valuable clinical explainability.
• Highlights potential for improved tumor segmentation and clinical decision-making through AI.