Measuring TB Cavities Honestly: Rater-Aware CT Burden as a Proxy Where Sequencing Is Missing

Most tuberculosis surveillance in high-burden regions never sequences the bacteria. A provincial hospital can count drug-resistant cases, but it cannot tell whether a patient caught an already-resistant strain from someone else or developed resistance during their own treatment. That distinction decides whether the response should be contact tracing or treatment support. Chest CT is often available where sequencing is not. Lung cavities, the holes that TB carves into tissue, are the classic marker of a highly infectious patient. So cavity burden measured from CT is an obvious stand-in for transmission risk. The catch is that cavities are exactly the structure radiologists disagree about. Their borders merge into the surrounding consolidation (solidified, inflamed lung). One benchmark gave up on separating the two and merged them into a single label.

This note proposes to treat that disagreement as part of the measurement rather than noise to be averaged away. We would collect multiple expert delineations of TB cavities and walls, then train a 3D model that reproduces the spread of expert opinion instead of a single consensus mask. We would then carry that spread through to the numbers surveillance actually uses: whether a cavity is present, how many there are, and their total volume. The central question is whether a cavity burden reported with a calibrated range ranks patients by genomically confirmed onward transmission better than a consensus-derived point estimate. A second question is equally important: this approach should not be expected to rescue drug-resistance prediction from CT. We pre-register that as a falsification test.

Core Claim

Imaging should be used for transmission-risk triage in the genomic surveillance gap, not for resistance calling. The quantity it offers, cavity burden, is currently unreproducible across readers. A probabilistic 3D segmentation model conditioned on reader style would make cavity burden reproducible and would attach honest uncertainty to it. That uncertainty should improve decisions such as which isolates to send for scarce sequencing and which index cases to trace first, compared with deterministic segmentation trained on consensus labels.

The contrast with prior work is specific, and each paper plays a different role.

Several papers motivate the problem. Epidemiology of drug–resistant tuberculosis in Hunan China over a 10–year period documents 6,597 drug-resistant cases characterized only by culture-based susceptibility testing. Its authors state that without genotyping they cannot separate acquired resistance from community transmission. Jung et al. (2024) hit the same wall for carbapenem-resistant bacteria in an intensive care unit.

Other work supplies the method. The Latent Doctor Model for Modeling Inter-Observer Variability models annotator subjectivity with a conditional variational autoencoder, a generator that samples plausible annotations from a learned latent "reader style." It builds that style code with a small set encoder over past annotations. We reuse this idea, but that work only handles 2D pathology grading, whereas we need 3D volumetric segmentation of a two-class boundary.

Other work constrains the design. Overview of the ImageCLEF 2017 Tuberculosis Task and Overview of ImageCLEFtuberculosis 2018 report that CT-only prediction of multi-drug resistance topped out at AUC 0.5825 and 0.6178. The organizers attribute this to resistance being a molecular trait without a reliable macroscopic signature. We take that seriously. It is why we aim imaging at transmission risk, where cavitation is a known driver, rather than at resistance.

DeepPulmoTB is the closest imaging work. It merged cavity walls and consolidation into one class because annotators could not separate them, and it trained a deterministic iterative-refinement network against a single consensus, reaching 69.6% Dice. We un-merge that class and model the disagreement explicitly.

Calibrate the Inter-Observer Segmentation Uncertainty via Diagnosis-First Principle (DiFF) reweights raters so the resulting mask best helps a downstream classifier. We deliberately do not do this for our main outcome. Tuning the boundary to a transmission label would make the burden measurement circular. We keep DiFF as a competitor on the downstream prediction tasks instead.

Sundermann et al. (2025) shows what genomic surveillance can do when sequencing exists: whole-genome sequencing plus electronic health record traces, with 91.7% adjusted sensitivity for transmission routes. Our work targets the setting where that pipeline is unaffordable.

Research Program

The first component is data, and most of it is new. We would select roughly 150 scans from the DeepPulmoTB / ImageCLEF CT pool, stratified by the ImageCLEF severity labels. Five thoracic radiologists would each delineate cavity lumen and cavity wall as separate classes. A 25-scan subset would be re-read after a washout period to estimate each reader's disagreement with themselves. This is the minimum set that makes reader-to-reader variability measurable on the exact structure DeepPulmoTB found ambiguous.

The second component is the model. A SwinUnetR backbone would be extended with a latent reader-style variable, following the Latent Doctor Model. It would come in two versions: one where the style prior is conditioned on the image, so ambiguous scans produce wider spread, and one conditioned on a few past annotations from a specific reader. Memory limits for 3D volumes are real. Uniformizing Techniques to Process CT Scans with 3D CNNs shows that spline-interpolated resampling works well for this data, so we would train on lung-cropped, uniformly resampled volumes.

The third component is propagation. For each scan we draw many mask samples and compute a distribution over three quantities: cavity presence, cavity count, and total cavity volume. The model is judged on whether those distributions cover a held-out reader's measurements at the nominal rate. Pixel overlap is a secondary criterion.

The fourth component is linkage, which requires a prospective partner cohort with CT and whole-genome sequencing. The candidate is a site like Hunan Chest Hospital, which already has CT and culture-based testing. We would sequence a retrospective sample of drug-resistant isolates and define genomic clusters by pairwise SNP distance, meaning the number of single-letter DNA differences between isolates. We then ask whether a patient's cavity-burden distribution predicts two things: membership in a cluster, and being the earliest-diagnosed member of a cluster. Finally, we simulate a fixed sequencing budget, allocated by cavity burden either with or without uncertainty.

Baselines

The naive reference point uses no imaging at all. Sequencing is allocated at random, or by demographic risk factors such as those in the Hunan cohort (farmer occupation, Western Hunan residence). Cluster membership is predicted from those same factors alone. Any imaging method must beat this, or imaging adds nothing to triage.

The strongest segmentation competitor is a deterministic network trained on consensus labels. We would retrain DPTBNet from DeepPulmoTB on majority-vote and STAPLE fusions of our five readers. STAPLE is a standard statistical method that estimates each reader's reliability and combines their masks. DPTBNet reports 69.6% Dice and 55.8% IoU against its merged consensus class. No result is reported on Lacuna for STAPLE or nnU-Net on TB cavity segmentation. It would be run with the same backbone family, the same resampling, the same training compute, and the same hyperparameter search budget as our model. Beating it means two things. First, higher coverage of held-out reader cavity volumes. Second, a higher agreement statistic (intraclass correlation, which measures how consistent repeated measurements are) between the model's burden and each held-out reader's burden. The comparison must also show no loss in Dice against that reader's mask.

For the downstream prediction tasks we would run DiFF, reweighting raters toward the downstream label. We would also run the strongest reported CT-only predictors. For severity, the reported results are AUC 0.7708 (MedGIFT) and RMSE 0.7840 (UIIP_BioMed) in ImageCLEF 2018, and 73% AUC for spline-interpolated zoom with a 3D CNN on ImageCLEF 2019 (Uniformizing Techniques). For drug resistance, the reported best is AUC 0.6178 in ImageCLEF 2018.

Two ablations isolate the key ingredient. One keeps our architecture but trains it on consensus masks with ensemble or test-time-augmentation uncertainty. This separates uncertainty about the model from disagreement among readers. The other keeps the reader distribution but merges cavity and wall back into C-LCW. This tests whether un-merging the classes matters.

Why not just…?

Why not just sequence every isolate and use whole-genome sequencing plus health-record mining, as EDS-HAT does? Where that is affordable it is the better tool, and we do not compete with it. Sundermann et al. (2025) reports 91.7% adjusted sensitivity in a tertiary US hospital with routine sequencing. The Hunan cohort shows the regime we target: a decade of cases with no genotyping at all. Our budget-constrained triage simulation measures directly how much of full-sequencing cluster detection an imaging-guided sample of, say, 20% of isolates recovers. If random 20% sampling recovers nearly as much, the imaging arm is not worth deploying, and we would say so.

Why not just predict severity or transmission risk directly from the CT volume and skip segmentation? This is a genuine threat. A 3D classifier already reaches 73% AUC on severity. Our answer is that a direct classifier produces a score no one can audit or compare across sites, whereas cavity count and volume are quantities clinicians already reason about. That answer only holds if the measured burden is at least as predictive. The linkage experiment therefore includes a direct CT-to-cluster-membership classifier with the same compute budget. If it clearly wins on held-out sites, the proposal narrows to a reproducibility contribution.

Why not just use majority vote, or an ensemble for uncertainty? Majority vote yields one mask. It therefore cannot express that a scan's cavity volume is genuinely ambiguous, and DiFF reports that it loses up to 2.73% diagnostic AUC against rater-aware methods. Ensembles are the stronger alternative. The Latent Doctor Model paper itself reports only parity with deep ensembles on one of its datasets. Our first ablation settles this: ensemble spread on consensus labels either covers held-out readers' measurements as well as the reader-conditioned model or it does not. If it does, we adopt the simpler method.

Why not just apply DiFF with drug resistance as the downstream target to break the "MDR wall"? This is the obvious intersection idea, and we expect it to fail. Resistance lacks a consistent macroscopic signature, per both ImageCLEF overviews. Tuning boundaries to a noisy label also invites overfitting on roughly 250 training patients. We run it as a falsification check rather than as the contribution.

Proposed Experiments

The decisive reproducibility check comes first. On the multi-reader subset, with scans held out by patient and one reader left out in turn, we compare the models on four measures. The first is intraclass correlation of cavity volume and count. The second is interval coverage at 80% and 95%. The third is the generalized energy distance between sampled and expert mask sets, which measures how closely the set of sampled masks matches the set of expert masks. The fourth is Spearman correlation between predicted spread and actual reader disagreement per scan, using 0.728 from the Latent Doctor Model on 2D pathology as a reference level rather than a target. Success means the reader-aware model's intervals achieve near-nominal coverage while the consensus baseline's point estimates fall outside readers' range on the ambiguous scans. Failure on coverage ends the propagation argument.

The downstream checks run on public data. We compare burden-derived features against the direct 3D CNN and DiFF on ImageCLEF 2018/2019 severity, where success is matching or exceeding 0.7708 AUC. We also run the pre-registered drug-resistance test on the ImageCLEF 2018 split. The prediction is that no variant exceeds 0.6178 AUC by more than noise across seeds. Confirming this strengthens the case that imaging belongs in transmission triage, not resistance calling. Refuting it would be a surprising and important result.

The linkage check is the one that makes or breaks the surveillance claim, and it needs new sequencing. We would sequence isolates from several hundred drug-resistant patients with CT at a partner site and define SNP-distance clusters. We then test whether uncertainty-aware cavity burden improves discrimination of cluster membership and of the likely index case, compared with consensus burden, demographic risk factors, and the direct classifier. We also test whether, under a fixed sequencing budget, allocating by the upper end of the burden interval recovers more clusters than allocating by the point estimate. If uncertainty adds nothing over the point estimate here, the honest conclusion is that reader variability matters for reproducibility but not for surveillance decisions.