Explainable AI
Making model decisions inspectable, which is a precondition for clinical use rather than an add-on. Evidenced by Grad-CAM in NeuroVision AI and RetinaVision AI, SHAP in the Parkinson's work, and "Explainable" in the title of the published paper.
A constraint on the other three directions rather than a project of its own. A model that reports a class and nothing else gives a clinician nothing to overrule, so each system carries the explanation path its modality allows: Grad-CAM over the brain MRI and retinal classifiers, SHAP over the voice-feature models for Parkinson's detection, input-gradient saliency for the echocardiography segmentation and 3D Grad-CAM for the coronary CTA volume. Flattening all of those to "Grad-CAM" would be convenient and wrong.
The harder half is the absence of evidence. A saliency map is produced for every prediction, including one the model had no basis for, so explainability also means being explicit about what is not there: a modality with no trained model returns no output at all rather than a plausible number, a segmentation below a minimum area is reported as not identified, and a Dice score measured on three held-out cases carries that fact in the same field as the number.
None of these explanations has been evaluated with clinicians. Whether they actually support a decision is an open question, and it stays an open question until it is tested with the people it is for.