Gao 2026 — TRICARE (deep-learning triage of 3D pathology)

A 2.5D MIL model that scores every depth level of a light-sheet-imaged biopsy and hands pathologists the riskiest few — genuinely externally validated on prostate, only internally on oesophagus, and with a reader study whose ‘upgrades’ have no ground truth behind them except a 59-case recurrence check.

A 2.5D MIL model that scores every depth level of a light-sheet-imaged biopsy and hands pathologists the riskiest few — genuinely externally validated on prostate, only internally on oesophagus, and with a reader study whose “upgrades” have no ground truth behind them except a 59-case recurrence check.

Read status: full text (publisher PDF) plus data records and code. Extended Data figures 1–8 and the supplementary tables/videos are referenced throughout and were not read; where a claim rests on them it is marked. Nothing was run.

What it is, in one paragraph for a reader who has not met 3D pathology

The specimen is not sectioned. A whole FFPE biopsy is dewaxed, stained in bulk with a nuclear dye (TO-PRO-3) and eosin, optically cleared (ethyl cinnamate) and imaged intact by an open-top light-sheet (OTLS) microscope — a fluorescence system that sweeps a thin sheet of light through the block and captures the whole volume at roughly 1 µm laterally, 3–4 µm axially. The two fluorescence channels are then false-coloured to look like H&E via a Beer–Lambert absorption model. The result is a stack of hundreds to thousands of virtual sections through the entire biopsy, where conventional histology cuts a few from one arbitrary side (the authors’ figure: under 1% of the volume). The reversible protocol leaves the block usable for IHC or molecular assays afterwards. The problem the paper sets itself is the obvious consequence: nobody can look at a thousand levels per biopsy, so which ones should the pathologist see?

The method

TRICARE is 2.5D multiple-instance learning. Each level is tiled (256 px ≈ 200 µm for prostate, 512 px for oesophagus), tiles are embedded by a frozen CONCH encoder (512-d; chosen for the low-data regime and because its training corpus includes special stains) through a small adapter layer that absorbs the H&E-to-synthetic-H&E domain shift, and then two attention stages run in a deliberate order: gated attention across tiles within a level (ABMIL) gives one vector per level, then non-gated attention across the target level and its neighbours gives a depth-aware vector that the classifier reads. The ablations are the interesting part: the winning configuration uses exactly one neighbour above and one below (m = 1), at ±60 µm for prostate and ±10 µm for oesophagus, and more or closer levels hurt — the authors read those distances as the scale of the diagnostic feature (glandular architecture versus cytology), which is a nice mechanistic claim. Lateral-then-depth beats depth-then-lateral and beats naive pooling of all patches from all levels; non-gated attention beats averaging, a linear layer and an RNN for the depth step. CTransPath as encoder reproduces the ordering (Extended Data 1, unread).

The fields that are usually buried

Prostate development	54 UW prostatectomy patients → 112 cancer-containing simulated 1×1×15 mm core biopsies; 121 labelled levels (53 low-grade GG1 / 68 GG≥2), 3,993 images with neighbours; majority of six GU pathologists on one or two central levels per biopsy
Prostate test	UPenn, 29 patients, one 3-mm punch each, 87 levels (69/18), majority of three; different staining protocol and a newer OTLS generation — signal, contrast, resolution and noise all differ
Oesophagus development	24 UW patients, 95 biopsy/EMR specimens, 334 levels (64 dysplasia/cancer / 270 benign), 6,978 images; majority of three GI pathologists
Oesophagus test	UW, 5 patients, 30 specimens, 77 levels (34/43) — same institution, same protocol, same microscope, held-out patients
Imaging	OTLS gen-2 at 0.44 µm/voxel (prostate dev), gen-4 at 0.37 (prostate test), gen-3 at 0.21 (oesophagus); all fused at 2× downsampling
Validation	Leave-one-patient-out CV on development cohorts; final model retrained on all development data with CV-chosen hyperparameters, then applied once to the test cohorts
External validation	Prostate: yes — different institution, protocol and instrument. Oesophagus: no — the “independent validation cohort” is internal

Results

	2D ABMIL AUC	TRICARE AUC
Prostate development (LOO-CV)	0.871	0.939 (p < 0.005)
Oesophagus development (LOO-CV)	0.895	0.921 (p < 0.005)
Prostate test (UPenn, external)	0.779	0.847 (p < 0.005)
Oesophagus test (UW, held-out)	0.882	0.917 (p < 0.01)

DeLong intervals throughout. The prostate drop from 0.939 to 0.847 across the institution and instrument change is the paper’s most informative number and the authors attribute it to exactly that.

Reader study (Fig. 4): 59 prostate and 30 oesophagus specimens imaged in 3D then sectioned conventionally; five (prostate) and three (GI) pathologists read TRICARE’s top three (prostate, ≥60 µm apart) or top eight (oesophagus, ≥20 µm apart) levels, and separately the standard 2D slides, one-month washout. Prostate: 8 cases upgraded to higher-grade on triaged 3D, 3 downgraded (with low reader agreement on the downgrades). Oesophagus: 5 upgraded, 0 downgraded, while reviewing eight images instead of sixteen sections. Fleiss’ κ was comparable between modalities — prostate 0.357 (3D) vs 0.415 (2D), oesophagus 0.770 vs 0.764. Only 11.9% of prostate and 13.3% of oesophagus samples had their model-predicted highest-risk level inside the depth window conventional sectioning would have sampled. For prostate, pathologist labels from each modality were used to predict 5-year biochemical recurrence: balanced accuracy 0.633 vs 0.614, F2 0.654 vs 0.588 in favour of triaged 3D.

Reading it critically

1. “Upgraded” is not “correct”. The headline clinical claim — better detection of high-risk disease — rests on pathologists calling more cases high-risk when shown the model’s chosen levels. Without an outcome, that is a shift in calls, not an improvement in accuracy; more levels shown could produce more upgrades by over-calling alone. The one outcome-anchored check is the recurrence analysis, and it is 59 cases with differences of 0.02 in balanced accuracy and 0.07 in F2, reported without an interval. That is compatible with no difference. The oesophagus arm has no outcome anchor at all.
2. The “<1% sampled / 88% of highest-risk levels lie outside the 2D window” argument is circular in its second half. “Highest-risk level” is the model’s prediction; using it to show that conventional sectioning misses the highest-risk level assumes the model is right about where risk is. The volumetric undersampling premise is true on geometry alone; the 88% figure is not evidence for it.
3. One external validation, not two. The oesophagus “independent validation cohort” is same-institution, same-protocol, same-microscope, held-out patients — that is a proper internal test set and it should be named as one. The prostate cohort is the real external test, and its 0.09 AUC drop is the transferability estimate this paper actually provides. See External Validation.
4. Labels at the agreement ceiling. Prostate training labels are the majority of six pathologists on one or two levels; the reader study’s own Fleiss’ κ for prostate is 0.36–0.42, i.e. fair agreement. The model is therefore trained and scored against a reference whose reliability the paper itself measures as modest — the ceiling Interobserver Agreement describes, and to the authors’ credit the number is reported, which most papers do not do.
5. Hyperparameters and design were selected on the same leave-one-out CV that produced the development AUCs. Epochs (50/100/150 by AUC), depth range, level count, encoder, aggregation order — all chosen on the development cohorts. The development-cohort numbers are therefore optimistic in the familiar way; the test-cohort numbers are the ones to quote, and the paper does keep them separate.
6. Small test sets, badly imbalanced. Prostate test: 18 higher-grade levels among 87, from 29 patients; oesophagus test: 5 patients. DeLong intervals are reported but the 5-patient oesophagus test is closer to a pilot than a validation.
7. The synthetic-H&E domain shift is handled by a single adapter layer and not otherwise examined. CONCH has never seen false-coloured light-sheet fluorescence; the paper’s own prostate-test drop mixes instrument, protocol and site effects with no way to apportion them. An encoder pretrained on OTLS data would be the obvious next experiment and is not attempted. (Reasonable at this stage; worth naming.)
8. Weighted sampling plus balanced accuracy and F2 at 90% sensitivity is sound handling of the oesophagus imbalance (64/270). The F2-at-fixed-sensitivity choice is well argued for a triage tool.
9. Competing interest, one line: the senior author co-founded and holds equity in Alpenglow Biosciences, which licenses the OTLS technology. Declared.

Why it earns space here

1. A third MIL variant for Weakly Supervised Learning, and the first one where the bag has an axis: instances are ordered in depth, and the design question — aggregate laterally first or across depth first, how many neighbours, how far — has answers that map onto tissue scale. Nothing in the group’s 2D MIL work has that structure, but the aggregation-order result is the kind of thing that transfers to serial-section or multi-level 2D problems.
2. A worked example of internal-versus-external validation labelled honestly on one arm and not the other, with the transferability drop measured — useful for External Validation precisely because both cases sit in one paper.
3. The reader-study design is the template and the cautionary tale at once: same specimens through both modalities, washout, multiple readers, κ reported — and still no ground truth for what an “upgrade” means. The group’s own reader-study instrument in QuPath Patoloji Atlası Extension would face the identical question.
4. The datasets are public and enormous, and the only 3D-pathology-with-outcomes data in the open; recorded on OTLS 3D Pathology Datasets with the honest note that the group has no instrument that produces this modality.

Related: OTLS 3D Pathology Datasets — the Zenodo and TCIA releases, sizes, licences, and what each actually contains. Related: Weakly Supervised Learning — 2.5D MIL as the depth-ordered variant beside classification and segmentation MIL. Related: External Validation — one genuine external cohort with a measured drop, one internal cohort called independent. Related: Interobserver Agreement — Fleiss’ κ reported for both modalities, and prostate labels sitting at κ ≈ 0.4. Related: Foundation Models in Pathology — CONCH as a frozen encoder pushed onto a modality it never saw, with a one-layer adapter as the whole domain-shift treatment. Related: Whole Slide Imaging — the substrate assumption this modality removes: no slide, no section, no scanner.

The repository (alecgao066/TRICARE, 6 commits ending 2026-08-12, single author, CC BY-NC 4.0) gets no separate project page — §6 gate, curvereg precedent. From the code: it is CLAM-derived (models/model_clam.py), with the depth-attention modules (Depth_Attn_Net, Depth_Attn_Linear) as the addition; create_splits_seq.py builds patient-stratified folds with a --leave_one_out flag, matching the Methods. Feature extraction is not in the repo — users are pointed at CONCH’s own repository — so it is a training harness over pre-extracted .pt features, not a pipeline.