Unrolling the Colon Wall: Depth-Preserving Canonical Coordinates for Interpretable Polyp Screening

Computer-aided polyp detection in CT colonography, the "virtual colonoscopy" that searches an abdominal CT scan for growths on the colon wall, already finds almost every large polyp. Its weakness is false alarms. In one clinical cohort, about 8,940 false candidates came from 162 studies that contained only 160 true polyps (Yao and Aman). Most false alarms are haustral folds (the ridges lining the colon), residual stool, or fluid. On any single view these look like polyps. A radiologist separates them by using context the detector lacks: whether the bump continues as a ridge around the colon, whether its tissue is continuous with the wall beneath it, and whether it stays put when the patient rolls from lying face-up (supine) to face-down (prone).

We propose to give a learned classifier that same context by changing the coordinate system rather than the network. The colon wall and the tissue behind it would be mapped, as a thick three-dimensional shell, onto a flat slab with three axes: distance along the colon, angle around it, and depth through the wall. The mapping is chosen to preserve angles, which is what "conformal" means. In these coordinates, folds become lines running across the colon, polyps become bumps with a measurable base, and the same spot of wall has the same coordinates in the supine and prone scans. The main payoff is interpretability. Clinically meaningful properties such as polyp height in millimetres, footprint area, how fold-like a candidate is, the density profile through the wall, and whether the candidate stays put between the two scans can be computed from the geometry instead of annotated by experts. That removes the annotation cost that normally limits "concept bottleneck" classifiers, which are models forced to predict human-readable intermediate properties before the final decision.

Core Claim

A small classifier that reads a candidate from this depth-preserving canonical slab, and must route its decision through five geometrically computed concepts, will reject more false positives at fixed sensitivity than the same classifier reading an ordinary three-dimensional voxel patch. Its explanations will also be verifiable, because every concept is a physical measurement. We further claim that the global map, not just local reorientation, accounts for a measurable share of the gain. The part of the gain we attribute to the global map comes from fold alignment and supine–prone correspondence, and an ablation is designed to separate the two.

Relation to Prior Work

Three lines of geometric work motivate the representation but stop short of learning with it. Zhu et al. and the companion SPIE paper conformally unroll vessel surfaces and paint curvature on them so humans can spot lesions. Su et al. turn cortical parameterizations into shape signatures for an SVM, reaching 87.5% accuracy for Alzheimer's versus controls. Nadeem et al. use the second eigenfunction of the surface Laplacian (the Fiedler vector, a smooth function whose extremes lie at the two ends of the colon) as a longitudinal coordinate. They use it to match supine and prone colons with a 5.24 mm error, against about 13.20 mm for centerline matching, and to detect folds. All three flatten only the two-dimensional surface, and none feeds a deep classifier.

We borrow Nadeem's longitudinal coordinate and fold detection directly, and add what surface methods cannot provide: the transmural dimension. For that we reuse the volumetric method of Abulnaga et al. and its journal extension. They map a tetrahedral placenta mesh to a flat template by minimising the symmetric Dirichlet energy $\|J\|_F^2+\|J^{-1}\|_F^2$, where $J$ is the local Jacobian of the map, with a line search that forbids inverted elements, and report 0.26% distortion. That work handled a disk-like organ and was never used for diagnosis. We apply it to a tubular shell, which introduces a seam and an elongated domain.

Theory from Gu's review of computational conformal geometry supports the choice of a cylinder as the canonical target. Zeng et al. show that distortion under discrete Ricci flow grows linearly with boundary perturbation. We rely on that result when arguing that imperfect lumen segmentation should degrade the map gracefully rather than catastrophically.

Interpretability work supplies the output design. Wu et al. show that a 1.5M-parameter 3D network predicting nine nodule attributes alongside malignancy reaches 97.58% accuracy on LIDC-IDRI, against 86.84% for a multi-crop CNN. Their attributes, however, needed radiologist labels. Pisov et al. show that replacing an opaque regressor with a head that predicts an explicit anatomical structure cuts midline-shift error from 2.92 mm to 1.54 mm and yields spatial confidence intervals. We follow Pisov in making outputs physical measurements and Wu in making them a bottleneck. Our difference is that the concepts come free from the geometry.

The volumetric-context literature explains why we do not simply use slices. Roth et al. had to smooth 2D predictions in 3D to raise pancreas Dice from 56% to 68%. Cheng et al. name missing 3D context as the main source of false positives in their 2D DeepLesion pipeline.

Research Program

The pipeline starts from a standard lumen segmentation. We offset the lumen surface outward along its normals to a fixed depth of roughly 10 mm, covering the wall and the adjacent fat, and tetrahedralise the resulting shell. The longitudinal coordinate $s$ is the Fiedler vector, following Nadeem. The circumferential coordinate $\theta$ is its harmonic conjugate on the lumen surface, which makes the surface map conformal onto a cylinder, with one cut routed between folds as Nadeem does. Depth $r$ is then extended through the shell by minimising the symmetric Dirichlet energy with the injectivity-preserving line search of Abulnaga et al. The lumen and outer boundary are pinned to $r=0$ and $r=R$.

Conformal maps preserve shape but not size. We therefore store the local area scale $e^{2u}$ (the factor by which the map stretches area at each point) and the volumetric Jacobian determinant as extra channels. CT intensities are resampled onto the slab so that an image grid is aligned with the colon's own axes. Supine and prone scans are mapped separately and aligned in $(s,\theta)$ using the fold anchors.

For each candidate produced by a fixed upstream detector, we extract a slab patch and compute five concepts. The first is protrusion height in millimetres, the lumen displacement above the local wall corrected by $e^{u}$. The second is base footprint area in square millimetres, the integral of $e^{2u}$ over the candidate base. The third is fold affinity, the elongation of the candidate along $\theta$ relative to $s$ and its distance to the nearest detected fold. The fourth is the transmural attenuation profile, whether soft-tissue density is continuous from the lesion into the wall or broken by bright tagged material. The fifth is supine–prone persistence, the displacement of the candidate between the two aligned scans.

A 3D CNN of Wu-scale capacity predicts these concepts from the slab patch. Training supervises each concept head with its geometrically computed value, so no manual concept labels are needed. The final polyp decision is a sparse linear model on the predicted concepts plus a small residual channel whose weight is penalised. We report how much of the decision flows through that residual.

Baselines

The naive reference is the upstream candidate detector with no false-positive reduction, which fixes the operating sensitivity and the starting false-positive count per case.

The classical geometric competitors are the curvature-and-texture system of Fiori et al. and the retrieval-based pruning of Yao and Aman. On the WRAMC database, Fiori et al. report 100% sensitivity with 0.9 false positives per case for polyps over 6 mm, and 93% sensitivity with 2.8 false positives per case for polyps over 3 mm. Yao and Aman report a 51.1% false-positive reduction at 97.5% sensitivity, and about 15% reduction at 100% sensitivity. Yao's retrieval system is the strongest reported number in the file for the exact task of pruning false candidates, and we treat it as the headline target. Where the data allow, we will re-run its $L_1$ retrieval over hand-crafted shape-index, curvature and Haralick texture features on our candidate set rather than compare across cohorts. Beating it means a larger false-positive reduction at 97.5% sensitivity, with non-overlapping patient-level bootstrap intervals, and a clearly larger reduction at 100% sensitivity, where retrieval is weakest.

The strongest deep competitor is a native 3D patch classifier on ordinary Cartesian voxels. It uses the same backbone as ours, with the same crop size in millimetres and the same augmentation, trained on the same patient-level folds with an identical hyperparameter search budget. It is also run with Wu-style multitask heads that predict the same five concepts computed in Cartesian space. For a deep 3D classifier on colonography false-positive reduction, we found no reported result on Lacuna, so this baseline must be run rather than quoted.

The key ablation replaces the global slab with local tangent-frame resampling. A patch is reoriented along the surface normal at each candidate, which gives the depth profile and rotation normalisation, but has no global longitudinal axis, no fold alignment and no supine–prone correspondence. A second ablation keeps the slab but removes the $e^{2u}$ and Jacobian channels. This tests whether uncorrected conformal distortion damages size-dependent decisions, which is the tension Su et al. raised. A third ablation drops the bottleneck and lets the network decide freely from the slab. This measures what interpretability costs or gains.

Why not just…?

Why not just train a 3D CNN on Cartesian candidate patches? This is the most serious threat. Wu et al. show that small 3D patch networks already work very well on lung nodules. Such a network sees the full 3D neighbourhood, so in principle it can learn fold shape and wall continuity. Our argument is that folds appear in every orientation in scanner coordinates, which forces the network to learn rotation invariance from a few hundred positives. Persistence across two scans is also invisible to a single-scan patch. Neither argument is proven. The matched-budget comparison in the baselines settles the question. If the Cartesian network with Cartesian concepts matches ours, the canonical coordinates are not worth their cost, and we would report that.

Why not just resample a local patch along the surface normal instead of computing a global map? This is cheaper, avoids the seam, and probably recovers much of the depth benefit. We therefore make it the central ablation rather than argue against it. Local reorientation cannot supply the persistence concept, and it supplies only a weak version of the fold-affinity concept, because "along the colon" is undefined locally. We expect any advantage of the global slab to concentrate in fold-type and stool-type false positives, and we will report errors broken down by false-positive category. If the local version matches the global one within confidence intervals even on those categories, we will drop the global parameterization and keep the geometric concept bottleneck on local patches as the contribution.

Why not just use flattened surface renderings, as in Nadeem et al., with a 2D network? Surface flattening already provides fold alignment and supine–prone correspondence. What it discards is the depth profile, the one cue that separates tagged stool from soft-tissue polyps. The surface-only variant of our slab, with depth fixed at the lumen, tests this directly. We predict it will fail specifically on stool false positives.

Why not just add Grad-CAM to a black-box classifier? Veltmeijer et al. found that unconstrained models exploit background artifacts, and a heatmap only shows that this is happening without preventing it. A heatmap also cannot say "this candidate is 4 mm tall." We will still compute saliency on the unconstrained ablation, to check whether it highlights the same regions our concepts measure.

Proposed Experiments

The main experiment uses candidates from the WRAMC colonography database of Fiori et al. (300 scans, 150 patients, with supine and prone pairs) and, if accessible, the 162-study cohort of Yao and Aman. We use patient-level cross-validation. The primary metric is false positives per case at 97.5% and 100% per-polyp sensitivity, reported separately for polyps over 6 mm and 3–6 mm. We test the parameterization itself on the 20-scan NIBIB cohort with expert fold labels. There we measure injectivity (the count of inverted tetrahedra), symmetric Dirichlet distortion, and supine–prone registration error against Nadeem's 5.24 mm.

We use several decisive checks. The first is the error breakdown by false-positive category across the Cartesian, local and global variants. The second tests size estimation: whether the height and footprint concepts predict colonoscopy-reported polyp size, and whether removing the $e^{2u}$ channel biases size in areas the map stretched or compressed. The third is a concept-intervention test. Replacing a predicted concept with its geometrically computed value should move the decision in the clinically expected direction, and the residual channel should carry a small share of the logit. The fourth is a stress test. We perturb the lumen segmentation and simulate collapsed segments, then track how the map and the false-positive rate degrade. This checks whether the linear-stability behaviour reported by Zeng et al. carries over. It also documents the failure mode Nadeem et al. note, where spectral coordinates break down in collapsed or obstructed segments.

The project succeeds if the global slab beats both the Cartesian network and Yao's retrieval at matched sensitivity, and the gain is concentrated where the geometry predicts it should be.