Multi-Organ Segmentation From Partially Labeled and Unaligned Multi-Modal MRI in Thyroid-Associated Orbitopathy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Ophthalmology

Background:

• Thyroid-associated orbitopathy (TAO) is an autoimmune disorder causing orbital disfigurement and vision loss.
• Quantitative assessment of TAO using multi-modal MRI is crucial but hindered by segmentation challenges.
• Existing methods lack comprehensive segmentation for multi-modal, unaligned TAO MRI data.

Purpose of the Study:

• To develop a novel method for automatic multi-organ segmentation in TAO using partially labeled, unaligned multi-modal MRI data.
• To introduce a cross-modal attentive self-training (CMAST) approach for improved TAO assessment.
• To enhance quantitative analysis of TAO through accurate MRI segmentation.

Main Methods:

• Proposed a cross-modal pseudo label self-training scheme to refine labels for comprehensive segmentation.
• Developed a learnable attentive fusion module for aggregating multi-modal knowledge without strict pixel-wise alignment.
• Incorporated prototypical contrastive learning loss for cross-modal feature alignment.

Main Results:

• The CMAST method achieved promising performance in comprehensive segmentation of TAO-affected organs on T1 and T1c MRI modalities.
• Demonstrated superior performance compared to previous methods on a large clinical TAO cohort (100 cases).
• Validated the effectiveness of cross-modal attention and self-training for segmentation accuracy.

Conclusions:

• CMAST offers a robust solution for multi-modal MRI segmentation in TAO, addressing limitations of existing techniques.
• The method facilitates quantitative assessment, potentially improving clinical management of TAO.
• The developed approach shows significant potential for advancing TAO research and diagnostics.