Physics-informed CV & neural inverse solvers
End-to-end deep-learning models trained against the light-transport forward model — spectral reflectance to absorption and scattering coefficients in sub-ms, with physically consistent predictions on-device.
Problem
Recovering optical properties — absorption and scattering coefficients — from spectral reflectance is an inverse problem. The classical route is to fit each measurement against a light-transport forward model, iterating until simulated reflectance matches what the camera saw. It works, but it is slow, and it does not fit inside the time and power budget of a handheld device that has to return an answer while the clinician is still holding it against the skin.
The harder constraint is consistency. A model that simply regresses coefficients from spectra can produce numbers that look plausible but violate the physics of how light actually moves through tissue. For a clinical readout like oxygenation, predictions that drift away from the forward model are worse than useless — they are confidently wrong.
Approach
I built end-to-end deep-learning models that invert the spectral-reflectance-to-coefficients mapping directly, and trained them against the light-transport forward model rather than against labels alone. Constraining training with the forward model keeps the network's outputs physically consistent: the predicted absorption and scattering have to reproduce the observed reflectance when pushed back through the physics.
The models are written in PyTorch and optimized for the GPU with CUDA, with the architecture and training redesigned around the structure of the inverse problem instead of treating it as a generic regression. That combination of algorithmic redesign and GPU optimization is where the speed came from — the forward model that used to run in a fitting loop becomes a training-time teacher, and inference collapses to a single forward pass.
What shipped
The result is spectral reflectance to absorption and scattering coefficients in sub-millisecond inference, running on-device, with oxygenation and tissue-structure predictions that stay physically consistent instead of drifting off the forward model.
Against the prior approach, the redesign delivered roughly a 60× computational speed-up and about a 30% accuracy gain — enough to move optical-property inversion from an offline step into the live capture loop.