MwoA auxiliary diagnosis via RSN-based 3D deep multiple instance learning with spatial attention mechanism

Abstract

Migraine without aura (MwoA) is the most typical migraine disease in the clinic, which is endangered to human health and challenging to diagnose. Developing the auxiliary diagnosis algorithms of MwoA based on functional connectivity (FC) changes from resting-state functional magnetic resonance imaging (rs-fMRI) is an important research domain. However, existing auxiliary diagnostic methods mainly adopt a seed-based correlation method to extract FC, which are easily affected by subjective factors. Moreover, those methods neglect the relationship between changes in FC and disease duration. In this paper, we report a weakly supervised learning method aiming to tackle those issues. We propose a resting-state brain network-based 3D deep multiple instance learning with spatial attention mechanism (R3D-DMILSAM) framework, where the patient-level label is allocated to the rs-fMRI data that view as multiple instances of a bag. R3D-DMILSAM uses the group information guided independent component analysis (GIG-ICA) to generate the subject-specific resting-state brain networks (RSNs). After that, the designed spatial attention-based 3D deep multiple instance learning (SA3D-DMIL) is trained to perform the diagnosis of MwoA. SA3D-DMIL can automatically generate several semantic deep instances and discovers abnormal RSNs using spatial attention mechanism. Extensive experimental results on the MwoA dataset show that R3D-DMILSAM achieves an overall accuracy of 88.80% and AUC of 94.70%. The visual network obtains high weight, which could be used as a potential biomarker for individualized diagnosis of MwoA.