Deep adaptive learning predicts and diagnoses CSVD-related cognitive decline using radiomics from T2-FLAIR: a multi-centre study
Lili Huang, Zhuoyuan Li, Xiaolei Zhu, Hui Zhao, Chenglu Mao, Zhihong Ke, Yuting Mo, Dan Yang, Yue Cheng, Ruomeng Qin, Zheqi Hu, Pengfei Shao, Ying Chen, Min Lou, K. K. He, Yun Xu
Abstract
Early identification of cerebral small vessel disease related cognitive impairment (CSVD-CI) is crucial for timely clinical intervention. We developed a Transformer-based deep learning model using white matter hyperintensity (WMH) radiomics features from T 2 -fluid-attenuated inversion recovery images to detect CSVD-CI. A total of 783 subjects (161 longitudinally followed) were enrolled from three centres for model development and external validation, using a domain adaptation strategy. The model achieved AUCs of 0.841 (training) and 0.859/0.749 (validation cohorts), outperforming conventional machine learning models. The gradient-weighted class activation mapping approach highlighted WMH textural features, particularly the logarithm-transformed gray level size zone matrix features, as key contributors. These features were significantly correlated with CSVD macro- and microstructural changes, mediated age-cognition relationships and predicted longitudinal cognitive decline. Our findings indicate that WMH radiomics features, reflecting CI-related biological changes in CSVD, combined with a Transformer-based deep learning model, constitute a feasible, automated, and non-invasive tool for CSVD-CI detection.