Detection and prediction of chemotherapy-related cognitive impairment using quantitative neuroimaging and innovative AI tools

Chemotherapy-related cognitive impairment (CRCI), coined "chemo-brain," is experienced by up to 75% of cancer patients and can degrade quality-of-life for years post-treatment. Yet it lacks objective diagnostic criteria or predictive markers. Information-rich quantitative neuroimaging may reveal relevant biomarkers which AI approaches can efficiently utilize for CRCI detection and prediction. Using 2,011 T1-weighted MRI brain scans from 1,349 patients, we explored several binary classifier models for two purposes. (1) To distinguish which patients have CRCI at the time of the scan using images obtained during or after chemotherapy (N=1,442). (2) To distinguish which patients will develop CRCI using pre-treatment scans (N=569) and comorbidity data. We improved upon a convolutional neural network (CNN) which previously showed strong performance in the classification of Alzheimer's disease, adding uncertainty estimation and out-of-distribution techniques. We explored manual feature extraction and connectivity analyses to add interpretability to the black box CNN. We evaluated performance using ROC and precision-recall analyses. This work demonstrates successful preliminary development of new, reliable, patient-specific informatics to improve long-term care of chemotherapy patients.