Healthcare AI & Medical Imaging
Deep learning for clinical imaging. Evidenced by the structural-MRI dementia classification manuscript under review at Neuroscience Informatics, and by the NeuroVision AI and RetinaVision AI systems.
Four modalities, four separate models: brain MRI, retinal fundus photographs, echocardiography and coronary CT angiography. Each has its own dataset, its own split and its own reported metrics rather than one architecture reused under four names.
The recurring problem is not the architecture. Clinical data arrives in the wrong orientation, at the wrong spacing, from the wrong scanner and from a population the training set never contained, and the datasets a small lab can obtain are small, imbalanced and expensive to label. So the decisions that carry the result are the ones around the model: patient- and case-level splits so no subject appears on both sides of an evaluation, class-balanced training where the labels are skewed, a decision threshold chosen on validation and never on test, and a presence rule that reports a structure as not identified instead of asserting a clinical absence.
The manuscript under review at Neuroscience Informatics applies attention-based multiple instance learning to 2D and 3D structural MRI for classification of vascular brain injury etiologic dementia. This direction is active, and none of it has been clinically validated or used in care.