Object detection & semantic segmentation
Detection and segmentation models that localize regions of interest and classify tissue in multispectral captures — gating downstream inversion so oxygenation and tissue-structure outputs land on the right pixels.
Problem
Optical-property inversion only means something if it runs on the right pixels. A multispectral capture of a limb or a wound contains skin, background, specular glare, and tissue types that behave very differently under near-infrared light. Inverting indiscriminately across the whole frame produces oxygenation and tissue-structure maps that are contaminated by everything that is not the tissue of interest.
So the pipeline needs a front end that understands the image before the physics runs — one that can localize the region of interest and separate tissue classes, and do it reliably enough to sit inside a device on the regulated path to clearance.
Approach
I developed object-detection and semantic-segmentation models for the imaging pipeline. Detection localizes the regions of interest in the frame; segmentation classifies tissue at the pixel level. Together they gate the downstream optical-property inversion, so the expensive physics only runs where it is valid and the outputs land on the correct anatomy.
The models operate on multispectral captures rather than ordinary RGB, which means the detection and segmentation heads have to make use of spectral structure the tissue carries, not just spatial texture. This front end became a fixed contract for everything downstream: oxygenation inference and tissue-structure analysis both depend on the masks it produces.
What shipped
The detection and segmentation stage ships as part of a clinician-facing imaging device, gating inversion so oxygenation and tissue-structure outputs are computed on the right pixels rather than the whole frame.
Because it sits inside a cleared device, the work carried the discipline that comes with it — the pipeline had to be validated, reproducible, and defensible, not just accurate on a held-out set.