Intravascular OCT drops closed-form fits: DiffOCT inverts one image for per-pixel attenuation and transfers to 29,269 clinical frames
Synopsis
Intravascular optical coherence tomography (OCT) images entangle tissue attenuation with the imaging system's own physics, so attenuation has been estimated with simplified closed-form inversions or with learned inverses tied to a specific scanner. DiffOCT writes the Beer-Lambert attenuation model as a differentiable forward model and runs gradient descent on a single B-scan to recover attenuation per pixel, with no training data or segmentation. Across 5,120 inversions on five lipid phantoms, recovered values rose monotonically with scatterer concentration and matched curve-fitting slopes; the same procedure ran unmodified on 29,269 clinical pullback frames from 128 patients, though the clinical data show geometric transfer only, not calibrated absolute values.
Interpretation
Each pixel's brightness in intravascular OCT is set jointly by local backscattering, cumulative attenuation along the path, the confocal gate and system gain, so the same tissue appears at different intensities with depth and attenuation cannot be read directly from the image. Earlier approaches fell into three groups: curve fitting a single attenuation over a manually segmented region, depth-resolved closed-form ratios of local intensity to remaining energy, and learned inverses fixed after training on simulated data. Across 5,120 B-scan inversions on five lipid phantoms imaged at two focal positions each, recovered attenuation rose monotonically with intralipid concentration and the slopes at both focal positions were comparable to the curve-fitting reference, while absolute values ran systematically low.
The method passes a candidate attenuation map through a differentiable OCT forward model that combines Beer-Lambert attenuation with confocal amplitude modulation, then updates the map by gradient descent to reduce the mismatch with the measured image. The forward model is a replaceable module: swapping in other scattering physics or recovering a different optical parameter leaves the optimization procedure unchanged, unlike closed-form estimators that need an analytic inverse or learned estimators bound to their training distribution. On 500 synthetic A-lines the forward model agreed with the analytical Beer-Lambert solution to residuals five orders of magnitude below the lowest signal, and on 500 synthetic B-scans the optimizer converged from scrambled initializations to near-perfect recovery.
As the number of coupled parameters grows, only gradient-based optimizers keep their accuracy, which is what makes per-pixel inversion tractable. Gradient-free methods lost accuracy as parameters rose from 25 to 5,000 across eight problem sizes, with Random Search and CMA-ES reaching MAPE of 30.32% and 100.91%. Adam and L-BFGS-B held mean MAPE below 0.1% at every size, with all four methods sharing the same forward model, bounds and eight paired initializations.
Perspective
The demonstrated scope covers synthetic A-lines and B-scans, five liquid lipid phantoms and five agar phantoms, and 29,269 intravascular OCT frames from 128 patients in the Disrupt CAD-I through CAD-IV cohorts; the clinical part shows only that the method transfers unmodified to clinical acquisition geometry, with recovered values reflecting relative tissue contrast rather than absolute quantities.
In the single-parameter forward model the attenuation coefficient controls both pixel brightness and cumulative attenuation, and adding a per-pixel backscattering parameter raises rendered PSNR without removing the ambiguity; recovered values change under aggressive perturbation of the initialization range, and no per-pixel attenuation reference standard exists in vivo.