This one is less of a single-topic explainer and more of a map of the explainers. If the other posts are about particular bottlenecks, this is the post about why those bottlenecks keep showing up in my work in the first place.
From the outside the projects can look scattered. Pathology classification, skin lesion segmentation, lung CT, denoising, reconstruction, deformation, surgical vision. Different modalities, different tools, different outputs. But the pattern underneath them is more stable than it first seems.
My projects look more connected now than they probably did while I was doing them.
The early work was more image-classification heavy. Pathology, blood smears, skin lesions, and related medical image tasks were good ways to learn what biomedical AI looks like outside clean benchmark assumptions.
Those projects taught me about dataset bias, explainability, class imbalance, staining variation, and why a high score does not automatically mean the model is using the right evidence.
That was the first layer: images are not just pixels. They come from a messy acquisition and clinical context.
Over time, I became more interested in outputs that had shape.
A segmentation mask is already a step in that direction. A mesh makes it more explicit. A 3D reconstruction makes it unavoidable. A deformation field makes it physical.
That shift changed the kinds of questions I cared about. Instead of only asking whether the model recognized the right thing, I started asking whether the output could be measured, inspected, registered, or used downstream.
Robotic surgery and image-guided intervention sharpen that interest because geometry is not just for analysis. It can guide action.
If the anatomy is reconstructed poorly, the guidance can be wrong. If the registration drifts, the overlay becomes misleading. If deformation is ignored, the preoperative model becomes stale. If the tool changes the anatomy and the model does not update, the system is quietly lying.
That is why surgical perception feels like the right convergence point for my interests. It combines imaging, geometry, motion, uncertainty, and real-time constraints.
The projects that taught me the most were usually the ones where the simple framing was not enough.
A classifier needed context. A segmentation needed a surface. A reconstruction needed a prior. A denoiser needed edge preservation. A surgical vision system needed deformation awareness. A benchmark metric needed failure-case inspection.
The pattern was always the same: the first version solved the obvious task, and the useful version had to solve the hidden one.
The technical thread is not a specific architecture. It is a way of framing problems.
I like problems where the image is incomplete, the anatomy has structure, and the output has to be useful beyond the model itself. That is why geometry keeps showing up. It forces the model to make claims about the physical world.
The next step is to make the thread tighter: sparse clinical imaging, physically plausible reconstruction, deformation-aware registration, and geometry that supports intervention.
That is the space I want to keep working in. Not medical AI as a generic classifier, and not robotics as pure control, but the perception layer between them.
The part that turns messy clinical images into anatomy a robot or clinician can actually reason about.