Abstract WMP20: Deep Learning Radiomics Identifies White Matter Hyperintensity-Related Cognitive Decline Based on T2-FLAIR

Introduction: Early identification of white matter hyperintensity-related cognitive impairment (WMH-CI) through normal MRI images is of great significance for early clinical intervention to reduce the occurrence of dementia. The aim of this study was to develop a generalizable and interpretable deep learning model for screening WMH-related CI using radiomic features (RFs) from T2 fluid-attenuated inversion recovery (T2-FLAIR) images. Methods: A total of 783 subjects were enrolled from three medical centres. RFs for WMH lesions were extracted from T2-FLAIR images in each subject. A deep learning model with a hierarchical transformer architecture was used to leverage all extracted RFs, instead of feature selection, to develop and cross-validate the diagnostic model for cognitive dysfunction. Unsupervised domain adaptation was used to address the cross-centre data inconsistencies without labelling the outer centre data. A gradient-weighted class activation mapping approach was adopted to identify the RFs that contributed most to the diagnosis. Subgroup analysis, correlation analysis and mediation analysis were performed to confirm the clinical relevance of the model. Besides, we also verify the clinical importance of the model by microstructural pathology of WMH detected by diffusion tensor imaging. Results: The deep learning model showed robust diagnostic power for WMH related CI, with an area under the receiver operating curve (AUC) of 0.841±0.016 in the development cohort. The prediction accuracy, sensitivity, specificity, precision and recall were 0.793 ± 0.108, 0.798 ± 0.021, 0.800 ± 0.065, 0.716±0.055 and 0.793 ± 0.108, respectively. The model generalized well across different subgroups. It also performed well in other two external verification cohorts, with an AUC of 0.859 and 0.749, respectively. The visual representation showed that the most important features were textural features, which were also significantly correlated with clinical assessment scale and diffusion parameters. Conclusions: This study presents a non-invasive imaging biomarker that can identify WMH-CI patients and is applicable to all levels of hospitals since only conventional MRI images are needed.