Light-activated cancer treatments, such as photodynamic therapy (a drug that kills cells only where light switches it on), succeed in mice and often stall in people. The usual explanation is depth: tumors in mice sit millimeters under the skin, while human targets lie centimeters deep behind fat, muscle and bone. Yet the evidence on how deep light actually reaches in a given person comes from a handful of idealized models. These include a homogeneous water tank, or one stylized four-layer head with textbook optical constants. Each is, in effect, a single case study of anatomy. Meanwhile, hospitals hold millions of routine CT scans taken for unrelated reasons. Open tools already turn those scans into accurate maps of skin, fat, muscle and organs. "Opportunistic screening" is the practice of mining such scans for a second purpose at no extra cost.
We propose to run physically accurate light-transport simulations on a large, diverse set of segmented routine CTs. The goal is to measure the population distribution of light dose reaching specific target organs, and to replace the single-phantom estimate with an estimate that has error bars. We then ask a sharper question that decides whether this effort is worth anything clinically. Can a cheap per-patient rule computed from a scan predict who is optically treatable? Such a rule might be skin-to-target depth, or a layered slab model using that patient's measured fat and muscle thickness. Or does treatability depend on full 3D anatomy, so that per-patient voxel simulation is needed?
Core Claim
Current estimates of transcutaneous light penetration rest on single geometries. The study in Monte Carlo-Based Optical Simulation of Optical Distribution in Deep Brain Tissues Using Sixteen Optical Sources traces 800 nm light through one scalp–skull–CSF–gray-matter model in the GPU Monte Carlo package mcxlab. It reports gray-matter fluence of 1.75–1.87% of the source under a collimated Gaussian beam, and lists tissue variability across individuals as its main unresolved limitation. Investigation of the Mechanisms of Radio-Dynamic Therapy motivates internal light generation by citing roughly 7 mm penetration for 630 nm light, and simulates only a water phantom. Both papers motivate this proposal: each exposes the gap without filling it.
Our claim has two parts. First, population-level feasibility rates can be computed reproducibly from routine CT archives. By feasibility rate we mean the fraction of real patients in whom a fixed light protocol delivers a set dose to a named target. These rates will differ materially from single-model predictions, and we will report the size of that difference. Second, we will establish empirically which level of anatomical detail is needed to predict feasibility per patient. If a per-patient layered slab model matches full voxel simulation, that finding is itself the deliverable: a screening rule that any Comp2Comp-style pipeline could compute in seconds. If the slab model fails, we will identify the anatomy where it fails.
Research Program
We build a pipeline in four steps. First, segment each CT into tissue classes. Second, assign each tissue literature optical properties at a chosen wavelength: absorption, scattering and anisotropy, the last describing how forward-directed each scattering event is. Third, place a source on the skin over a named target. Fourth, run voxelized Monte Carlo photon transport in mcxlab, the tool the brain-model study treats as the reference method.
For segmentation we reuse existing work rather than invent new models. Comp2Comp: Open-Source Body Composition Assessment on Computed Tomography provides muscle, subcutaneous fat and visceral fat, with internal Dice of 0.97, 0.97 and 0.96 respectively and external Dice above 0.80–0.85. Dice is a 0–1 overlap score between predicted and true masks. TotalSegmentator, through Comp2Comp and Segmenting Whole-Body MRI and CT for Multiorgan Anatomic Structure Delineation, provides organs and bone; the latter reports liver Dice up to 0.96 and kidney Dice up to 0.95. Skin is taken as the outer body-contour shell.
We fix three target organs that TotalSegmentator segments reliably and that span shallow to deep: the thyroid, the anterior liver surface, and the kidney through the flank. We fix three wavelengths: 630 nm (red), 800 nm (first near-infrared window) and 1064 nm (second near-infrared window). The beam profile is the collimated Gaussian that the brain study found penetrates deepest. Optical coefficients are not known per patient. We therefore draw them from literature ranges, perturbing around the values used in the brain study, and propagate them through every simulation. As a result, each patient's dose is a distribution rather than a point.
The single-case lens becomes a measurement. We bootstrap the cohort to show how far a feasibility estimate drawn from one patient, or from five patients, or from the stylized phantom, lands from the full-population rate. This quantifies the "proof of possibility versus proof of pattern" gap that case reports such as Teru Hideshima et al. and Jiayu Zhou et al. acknowledge only qualitatively.
A secondary question concerns contrast agents. Iodinated contrast inflates CT intensity values (Hounsfield units, the CT brightness scale), which could flip voxels across a fat/muscle threshold or bias any intensity-to-blood-content mapping. Andrea Toelly et al. showed this bias is linear and correctable for bone density. Sebastian Rühling et al. detect the contrast phase from images at 98.3% internal and 94.2% external accuracy. We reuse their phase detector to test whether contrast phase changes simulated dose enough to matter. We expect the effect to be small, because tissue labels come from learned segmentation rather than intensity thresholds, but the experiment checks rather than assumes this.
Baselines
The naive reference point is the single-geometry estimate that the field uses now. For each target and wavelength, we build one stylized planar layered phantom with median textbook layer thicknesses and simulate it in mcxlab. Its single feasible/infeasible verdict is then applied to every patient. This reproduces the practice of Monte Carlo-Based Optical Simulation of Optical Distribution in Deep Brain Tissues Using Sixteen Optical Sources and of the water-phantom analysis in Investigation of the Mechanisms of Radio-Dynamic Therapy. We beat it if its predicted population feasibility rate falls outside the bootstrap 95% interval of the voxel-simulated rate for at least one target–wavelength pair. We also beat it if its per-patient accuracy is near the base rate, meaning it carries no individual information.
The second baseline is a skin-to-target depth cutoff, the caliper rule a clinician would actually use. We compute minimum skin-to-target distance from the same segmentations and pick the threshold on a training split to maximize agreement with voxel-simulated feasibility. It is a strong competitor because depth is the dominant variable in every penetration statement in the file.
The third, and strongest, baseline is a per-patient layered slab model. It takes that patient's measured skin, fat, muscle and intervening-organ thicknesses along the beam axis, assembles them into a 1D stack, and simulates that stack in mcxlab with the same photon count and the same coefficient draws as the voxel run. This isolates the one ingredient the proposal adds: full 3D geometry, including curvature, oblique layers and lateral heterogeneity. All three predictors are tuned on the same training patients, evaluated on the same held-out patients, and given the same coefficient uncertainty.
Beating the slab model means a statistically significant gain on held-out patients, by paired bootstrap, in two respects. One is the area under the curve (AUC) for classifying voxel-simulated feasibility. The other is rank correlation with target dose, for at least one target. The thyroid and flank kidney are the most likely candidates, because anatomy there is curved and non-planar. For population-level fluence prediction across real patients we found no reported result on Lacuna, so there is no published number to beat. The closest reported quantity is the single-model gray-matter fluence of 1.75–1.87% at 800 nm, which serves only as a sanity check for our head-free targets' order of magnitude.
Why not just…?
Why not just use skin-to-target depth? Depth may explain most of the variance, and our experiment measures exactly how much. The brain-model study gives a reason to doubt depth alone: the scalp absorbed about 69–71% of energy and CSF under 1%, so equal millimeters of different tissue attenuate very differently. A thick fat layer and a thick muscle layer at the same depth should therefore diverge. If the depth rule's AUC is within the bootstrap interval of the voxel model, we will say so and recommend depth.
Why not just run a per-patient layered model, or the diffusion approximation on it? This is the real threat, and it is the comparison the proposal is organized around. The diffusion approximation is a closed-form shortcut that treats light as spreading like heat. It is known to fail near boundaries and in shallow, weakly scattering layers, which is why we use slab Monte Carlo as the fair version of this alternative instead. If the slab model matches voxel simulation everywhere, the 3D component is unnecessary. The useful output then becomes a validated slab-based screening rule, which is cheaper and easier to deploy. We regard that as an acceptable and publishable result, not a failure.
Why not just measure tissue directly with ultrasound or diffuse optical tomography before treatment? Direct measurement is better for the patient already on the table. It cannot tell a trial designer, before anyone is scanned, what fraction of a population a protocol could reach, or flag eligible patients from scans that already exist. The two approaches are complementary. The proposal's population claims do not depend on direct measurement, while per-patient clinical use eventually would.
Proposed Experiments and Decisive Checks
The cohort combines public TotalSegmentator CTs with VerSe: A Vertebrae Labelling and Segmentation Benchmark for Multi-detector CT Images. VerSe contributes 374 scans from 355 patients across manufacturers and supplies scanner and anatomical diversity. Patients are stratified by subcutaneous fat area from Comp2Comp so that results can be reported across body habitus. We hold out patients for evaluating the predictors.
The first decisive check compares the population feasibility rate with the single-phantom verdict, together with the bootstrap curve showing how many patients are needed before the rate's 95% interval narrows below ten percentage points. The second decisive check compares the three per-patient predictors against voxel simulation, as specified above.
Robustness checks address three questions. First, do results survive the optical-coefficient uncertainty, or is between-patient variance swamped by between-coefficient variance? If the latter, the honest conclusion is that tissue optics, not anatomy, is the bottleneck. Second, do segmentation errors matter? We test this by perturbing masks at the reported Dice levels. Third, does contrast phase matter? We compare contrast-phase groups identified by the Rühling detector.
Where an archive has repeat scans of the same patient, we test whether per-patient feasibility is stable over time. This mirrors Intraindividual Reliability of Opportunistic Computed Tomography–Assessed Adiposity and Skeletal Muscle Among Breast Cancer Patients, which found body-composition intraclass correlations (a test–retest agreement score) of 0.763–0.998 across repeat CTs. Stable feasibility is what would make a scan from last year usable for a treatment decision this year.
The central limitation is that all ground truth here is simulated. Clinical validity would ultimately require measured fluence in vivo. This work aims to establish whether that expensive validation should target slab-level or voxel-level models, and for which organs.
References
- Monte Carlo-Based Optical Simulation of Optical Distribution in Deep Brain Tissues Using Sixteen Optical Sources
- Investigation of the Mechanisms of Radio-Dynamic Therapy
- Comp2Comp: Open-Source Body Composition Assessment on Computed Tomography
- Segmenting Whole-Body MRI and CT for Multiorgan Anatomic Structure Delineation
- A RESECTED CASE OF ADVANCED ESOPHAGEAL CARCINOMA SHOWING A PROMINENT RESPONSE TO NEOADJUVANT CHEMOTHERAPY WITH 5-FU AND LEUCOVORIN
- Laparoscopic pancreaticoduodenectomy in A-92-older Chinese patient for cancer of head of the pancreas
- Influence of Contrast Media on Bone Mineral Density (BMD) Measurements from Routine Contrast-Enhanced MDCT Datasets using a Phantom-less BMD Measurement Tool
- Automated detection of the contrast phase in MDCT by an artificial neural network improves the accuracy of opportunistic bone mineral density measurements
- VerSe: A Vertebrae labelling and segmentation benchmark for multi-detector CT images
- Intraindividual Reliability of Opportunistic Computed Tomography–Assessed Adiposity and Skeletal Muscle Among Breast Cancer Patients