Deep adaptive learning predicts and diagnoses CSVD-related cognitive decline using radiomics from T₂-FLAIR: a...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Neurology

Background:

• Cerebral small vessel disease (CSVD) is a common cause of cognitive impairment.
• Early identification of CSVD-related cognitive impairment (CSVD-CI) is critical for effective clinical management.
• Current diagnostic methods may lack sensitivity for early-stage CSVD-CI.

Purpose of the Study:

• To develop and validate a Transformer-based deep learning model for detecting CSVD-CI.
• To utilize white matter hyperintensity (WMH) radiomics features for CSVD-CI detection.
• To investigate the key WMH features contributing to CSVD-CI detection.

Main Methods:

• Development of a Transformer deep learning model using WMH radiomics from T2-FLAIR images.
• Training and external validation on a cohort of 783 subjects across three centers, employing domain adaptation.
• Utilizing gradient-weighted class activation mapping (Grad-CAM) to identify important radiomic features.

Main Results:

• The Transformer model achieved high performance with AUCs of 0.841 (training) and 0.859/0.749 (validation).
• The model outperformed conventional machine learning approaches in detecting CSVD-CI.
• WMH textural features, specifically gray level size zone matrix features, were identified as key contributors.

Conclusions:

• WMH radiomics features, analyzed by a Transformer model, offer a feasible and automated approach for non-invasive CSVD-CI detection.
• These radiomic features reflect biological changes associated with cognitive impairment in CSVD.
• The developed tool has the potential to aid in the early identification and management of CSVD-CI.