Research
Researching practical problems at the intersection of AI, machine learning, medical imaging, and intelligent healthcare systems.
Work in progress, with results either published or under review.
Developing machine learning and deep learning methods for medical image analysis, clinical decision support, and explainable healthcare AI.
Graph and transformer architectures applied to biological interaction data. Evidenced by X-GNN, published at IEEE ICECTE 2026, on human protein–protein interaction prediction.
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.
Dataset quality as the object of study rather than a preliminary step. Evidenced by the Data Annotation Analyst role at Rooya AI, the CVAT annotation portfolio, and dataset analysis work at AMIRL.
Being read into and prototyped. No results yet.
Exploring machine learning and computational modeling for neuroimaging and neuroscience, with an interest in understanding neural representations underlying cognition and emotion.
Open questions
Accuracy is not adoption. A model that cannot show why it reached a conclusion gives a clinician nothing to overrule, so the interesting question is not how to raise a score but how to expose the evidence behind one prediction at a time — and how to make the absence of evidence equally visible.
Imaging, signals and structured records disagree, and a system that averages them hides the disagreement. The question is how to fuse modalities in a way that is auditable — where a reader can still see which modality carried a conclusion, and where an untrained component contributes nothing rather than a plausible number.
Clinical images arrive in the wrong orientation, at the wrong spacing, from the wrong scanner, and with populations the training set never contained. The question is how much of that gap can be closed by data and evaluation design rather than by a larger model, and how to report honestly when a test split is too small to support a claim.
Foundation models are trained at a scale no individual researcher can reproduce, and medical datasets are small, imbalanced and expensive to label. The question is which adaptation strategies — transfer learning, parameter-efficient fine-tuning, careful supervision — earn their cost at that scale, and which merely look like progress.
Research methods
A research workflow that connects real-world problems, evidence, experimentation, modeling, evaluation, interpretation, validation, and reproducibility.
Problem Formulation
Translate a real-world problem into a precise research problem, objective, and measurable outcome.
Diagnostic trade-off
Increasing sensitivity can reduce false negatives, while increasing specificity can reduce false positives. The appropriate operating point depends on the consequences of each error.
A useful research result should survive realistic constraints rather than only perform well on a benchmark. The case study connects model behavior to the actual environment in which the method is intended to operate.
Research principle
The goal is not to use the most complex model.
It is to build the most appropriate, measurable, interpretable, efficient, and reproducible solution for the question.
