Multi-Task Weakly-Supervised Attention Network for Dementia Status Estimation With Structural MRI.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Neurology

Background:

• Predicting dementia progression using structural MRI is crucial for understanding Alzheimer's disease (AD) pathology.
• Current machine/deep learning methods often preselect brain regions, leading to heterogeneity and reduced performance.
• Existing approaches rely on anatomical knowledge and time-consuming registration, overlooking individual-specific brain changes.

Purpose of the Study:

• To propose a novel multi-task weakly-supervised attention network (MWAN) for joint clinical score regression from baseline MRI scans.
• To develop an end-to-end deep learning model integrating dementia-aware feature learning and multitask regression.
• To automatically identify subject-specific discriminative brain locations for improved prediction.

Main Methods:

• A backbone fully convolutional network extracts MRI features.
• A weakly supervised dementia attention block identifies subject-specific discriminative brain locations.
• An attention-aware multitask regression block jointly predicts multiple clinical scores (MMSE, CDRSB, ADAS-Cog).

Main Results:

• The MWAN method demonstrated superior regression performance compared to state-of-the-art methods on two public AD datasets.
• Quantitative experiments validated the model's effectiveness in predicting clinical scores.
• Qualitative results showed that automatically identified brain locations are biologically meaningful and retain individual specificities.

Conclusions:

• The proposed MWAN offers an effective, end-to-end deep learning framework for dementia progression prediction using structural MRI.
• The method overcomes limitations of traditional approaches by learning holistic features and incorporating individual-specific brain changes.
• MWAN's ability to identify biologically meaningful, subject-specific brain regions advances the understanding and prediction of Alzheimer's disease.