A dual-module 3D fusion framework for multi-modal MRI segmentation in anal fistulae

Recent advancements in deep learning have greatly impacted medical image segmentation, especially in the segmentation of human organs and tissues. Accurate segmentation is crucial for precise diagnoses and effective treatment planning. However, segmenting anal fistulae in magnetic resonance imaging (MRI) images presents a significant challenge due to the similarity between lesions and normal tissues, which often results in high false positive rates.

This study aims to develop a robust segmentation method for anal fistulae in MRI images that improves accuracy by reducing false positives, while also efficiently handling the complexity of multi-modal and multi-directional MRI data.

We propose a novel dual-module framework. At the input stage, a inspired method is employed to efficiently capture multi-dimensional information from the MRI data. At the output stage, residual techniques are applied to optimize and enhance the segmentation results, ensuring comprehensive integration of multi-modal and multi-directional MRI information. Additionally, an Outlier-Penalized Dice Loss (OPDL) function is introduced to specifically address the issue of false positives in segmentation.

The proposed method was evaluated on a dataset of 950 self-collected multi-directional MRI scans. Our approach demonstrated superior segmentation performance compared to six state-of-the-art methods, achieving a Dice Score of 0.7324 and an IoU Score of 0.5943, which represents significant improvements over the highest-performing baseline models.

The dual-module framework, combined with the selection mechanism and residual techniques, along with the innovative OPDL, offers a significant advancement in the segmentation of anal fistulae in MRI images. This approach not only improves accuracy by effectively distinguishing between lesions and normal tissues but also reduces computational complexity, making it a promising tool for clinical applications.