Null-Space Conformal Uncertainty for Deep Diffuse Optical Reconstruction

Near-infrared light that enters the head or breast scatters thousands of times before it reaches a detector. The measurements therefore carry less and less information about tissue as depth increases, and below some depth they carry almost none. Neural networks now reconstruct absorption maps from such measurements far better than classical solvers. When the data say nothing about a deep region, though, the network still outputs a confident answer, and that answer comes from its training prior rather than from the light. Standard uncertainty methods tend to miss these cases. Ensembles of networks trained on the same prior agree with one another in exactly the regions the data cannot see. Physics-residual checks also pass there, because a wrong deep structure can fit the boundary measurements perfectly.

We propose to split each reconstruction error into two parts using the physics of light transport. The first part is the component the measurements can detect. The second is the component they cannot, which lies in the "null space" of the linearized forward model: tissue changes that leave the detector readings essentially unchanged. We then build distribution-free error bars (conformal prediction) separately for each part, calibrated depth by depth. The central claim is testable. As depth increases, an increasing share of the true error of deep diffuse optical tomography (DOT) and functional near-infrared spectroscopy (fNIRS) reconstructions lies in this physics-defined null space. Uncertainty estimates built from network samples or from physics residuals undercover there. Bands built on a basis taken from the forward model's sensitivity matrix restore coverage at each depth without inflating band width near the surface.

Why this gap is real

The physics sets the problem. In tissue, scattering dominates absorption, and beyond about a millimetre light diffuses. Reconstruction is a severely ill-posed inversion of the diffusion equation $-\nabla\cdot(D\nabla\Phi)+\mu_a\Phi=S$, where $\Phi$ is photon fluence, $D$ is the diffusion coefficient, $\mu_a$ is absorption and $S$ is the source. Non-invasive fNIRS reaches only a few millimetres into cortex, and every photon passes through scalp blood flow twice.

Four learned DOT solvers improve point accuracy without addressing what the data cannot determine. Deep Learning Diffuse Optical Tomography learns a framelet inversion. DOTnet 2.0 denoises inside a Jacobian-driven loop and reaches MS-SSIM 0.80–0.90 against 0.36–0.59 for Tikhonov reconstruction. Unrolled-DOT unrolls a shrinkage-thresholding solver and reports SSIM 0.88 against 0.60 for ADMM/FISTA on 5,000 real time-of-flight scans. Mod-DOT and the GNN-prior method of Learnable Priors Support Reconstruction in Diffuse Optical Tomography compress signals into learned latent priors. All four output deterministic maps. Their strong priors are precisely what fills the null space, so their accuracy gains motivate this proposal: the better the prior, the more convincing an unsupported deep structure becomes.

Three uncertainty-quantification papers supply our machinery, and each has a limitation that shapes the design. Calibrated Physics-Informed Uncertainty Quantification (CP-PRE) uses PDE residuals as label-free conformal scores. It kept 90–95% coverage on Navier–Stokes under shift, where MC dropout and ensembles fell to 30–40%. For an inverse problem, however, a residual measures data consistency, and by definition it cannot see null-space error. CP-PRE is therefore our method for one half of the error and a negative constraint on the other half. CONSIGN builds spatially coherent conformal bands along the leading SVD directions of network samples. We keep its low-rank band construction but take the basis from the forward model's sensitivity matrix rather than from the samples, because samples cannot explore directions in which every network is wrong in the same way. Uncertainty-aware performance assessment of optical imaging modalities with invertible neural networks showed that, for surface multispectral reflectance, some tissue parameters such as blood volume fraction are fundamentally ambiguous. We read that as evidence that optical ambiguity is structured and physically predictable, and we extend the question from planar reflectance to depth. Finally, A systematic evaluation of uncertainty quantification techniques in deep learning: a case study in photoplethysmography signal analysis found that global calibration metrics hide local failure. That result is why we evaluate coverage conditional on depth rather than pooled across the volume.

Core Claim

Let $J$ be the sensitivity matrix of the measurements with respect to absorption, linearized at the network's estimate $\hat\mu$. Write $P_R$ for the projector onto right singular vectors of $J$ whose singular values exceed the measurement-noise floor, and $P_N = I-P_R$ for the projector onto the remaining, data-invisible directions. Any error $e=\mu-\hat\mu$ splits as

$$e = P_R e + P_N e .$$

The first component shows up in the measurement residual, so it can be calibrated without labels in the style of CP-PRE. The second component is invisible to the detectors. Only labeled calibration data drawn from the same anatomical distribution can bound it, so it is honestly epistemic. Within each depth slab $d$, we use the nonconformity score

$$s_d = \max\left(\frac{\lVert P_R e\rVert_d}{\hat\sigma_{R,d}},\ \frac{\lVert P_N e\rVert_d}{\hat\sigma_{N,d}}\right),$$

where $\hat\sigma_{R,d}$ and $\hat\sigma_{N,d}$ are per-slab scale estimates. Split-conformal quantiles of $s_d$ then give a band for each depth slab, built along the leading directions of each subspace. Because $J$'s singular values shrink with depth, $P_N$ automatically takes up more of each deeper slab. The band therefore widens where the light provides no information, not wherever a network happens to be unstable.

Research Program

We will build the decomposition on top of an existing learned solver rather than proposing a new reconstructor, so that any gain is attributable to the uncertainty layer. Simulation will use NIRFAST finite-element meshes in two families. The first family consists of slab and breast phantoms, using DOTnet 2.0-style vascular targets and Mod-DOT-style inclusions. The second family consists of four-layer head models (scalp, skull, cerebrospinal fluid, cortex) whose layer thicknesses vary across simulated subjects and which include superimposed scalp hemodynamic fluctuations. The second family is the fNIRS regime. In it, the most clinically important question is whether a reconstructed cortical change is supported by the light or merely attributed to cortex by the prior, when it may in fact be scalp signal.

On each family we will train a Mod-DOT-style solver and an Unrolled-DOT-style solver. We compute $J$ at the estimate with the finite-element adjoint, calibrate on held-out simulated subjects, and test under two conditions. The first is exchangeable data. The second is shift: unseen skull thickness ranges and unseen scalp-flow amplitudes. Shift breaks the exchangeability that conformal guarantees rely on. For that case we will compare a naive calibration with calibration stratified by an observable proxy, such as source–detector distance or a short-channel scalp estimate. We will report honestly how much of the coverage guarantee survives.

As a stress case, we will apply the same analysis to Balters-style cortical-to-subcortical inference (Towards Assessing Subcortical "Deep Brain" Biomarkers of PTSD with Functional Near-Infrared Spectroscopy, reporting $r>0.75$–$0.79$ from simulated fNIRS). Under our decomposition, amygdala activity lies entirely in $P_N$. Its uncertainty is therefore wholly epistemic, and it should be calibrated across subjects and groups, never certified by data fit. We cannot reproduce their fMRI cohort from the file, so this case stays in simulation with synthetic connectivity, and we flag it as a demonstration rather than a clinical claim.

Baselines

The do-nothing reference is the point-estimate solver with no uncertainty, together with naive voxelwise intervals equal to ±2 standard deviations of a five-member ensemble. It shows how badly unadjusted estimates undercover at depth. The strongest classical competitor is linearized Bayesian inversion: Tikhonov regularization read as a Gaussian prior, giving a closed-form posterior covariance that is already depth-dependent. We run it on identical meshes and noise levels, with its prior covariance fitted to the same training phantoms the networks use.

The strongest learned-uncertainty competitors are deep ensembles and MC dropout implemented in Torch-Uncertainty, plus CONSIGN applied to the ensemble samples. All three get the same number of forward passes as our method needs for the Jacobian computation, and all share our tuning budget and calibration split. We also adapt CP-PRE, using measurement residuals alone as the score. The file reports no coverage numbers for any of these on DOT ("no reported result found on Lacuna for DOT coverage"). The nearest reported reference is CP-PRE's maintained 90–95% coverage against 30–40% for dropout and ensembles on Navier–Stokes under shift. For point accuracy, the reference is Unrolled-DOT's SSIM of 0.88.

The key ablation swaps the basis. We keep everything else fixed and replace $J$'s singular vectors with the SVD of ensemble samples, which is CONSIGN's basis. A second ablation removes depth stratification. Beating the baselines means matching the nominal coverage, for example 90% ± 2%, in every depth slab including the deepest. Every baseline except linear Bayes should fail in the deepest slabs. It also means producing bands no wider than linear Bayes near the surface and narrower than linear Bayes wherever the learned prior is genuinely informative.

Why not just…?

Why not just use the linear Gaussian posterior, which already knows about the null space? This is the most serious threat. If the training prior were Gaussian, linear Bayes would be the right answer. Vascular trees and sparse inclusions are not Gaussian, though, and a Gaussian prior's null-space variance will be wrong in both directions: too wide where the learned prior is sharp, too narrow on outlying anatomy. The comparison of per-slab band width at matched coverage settles this. If linear Bayes matches us, we will report that the learned solver adds accuracy but no uncertainty benefit.

Why not just run split conformal on ensemble spread, binned by depth? Depth-binned calibration can repair the bin-level marginal coverage of almost any score, so this baseline could reach nominal coverage. It does so by inflating bands uniformly across the slab, including in data-determined directions. The basis-swap ablation measures exactly this difference, as band volume at equal coverage. If the volume gap is negligible, the physics basis is unnecessary and we will say so.

Why not just use CP-PRE's physics residuals, since they need no labels? A wrong deep inclusion can produce a near-zero measurement residual. Residual scores are therefore blind by construction to the component that grows with depth. Our experiments test this by reporting residual-only coverage per slab. We predict it decays with depth while remaining fine near the surface.

Decisive Checks

The idea stands or falls on three measurements. First, the fraction of true error energy lying in $P_N$ should rise monotonically with depth on both phantom families. Second, ensemble and CONSIGN bands should undercover specifically in that component, while our bands hold nominal coverage per slab. Third, under the skull-thickness shift, stratified calibration should recover most of the lost coverage. If the null-space fraction stays flat with depth, or if depth-binned ensembles match our coverage at equal band volume, the central claim is false, and the work reduces to a recommendation to use simple Mondrian calibration for diffuse optics.