QLabelMIL

Inter-Pathology Query Decoding for Multi-Label Gastric Histopathology

Pedro C. Neto^1,3,* · Rita N. Lopes^3,* · Lígia Prado e Castro²

¹ Unilabs.AI, Campus Biotech, Geneva, Switzerland · ² Unilabs, Campus Biotech, Geneva, Switzerland ³ Faculdade de Engenharia da Universidade do Porto (FEUP), Portugal · ^* Equal contribution

MICCAI 2026 · COMPAYL Workshop

---

TL;DR

Gastric cancer precursors (H. pylori infection, atrophy, intestinal metaplasia) co-occur along the Correa cascade, but most whole-slide image (WSI) classifiers train one aggregator per condition or share a single slide embedding. QLabelMIL is a multiple-instance learning (MIL) aggregator that:

• uses one learnable query per pathology and a transformer decoder to model patch-to-label relationships;
• supports joint multi-label classification in a single model;
• produces per-class attention heatmaps natively;
• optionally uses a Graph Convolutional Network (GCN) over the label co-occurrence prior to couple the per-label classifiers.

Evaluated on 5,764 WSIs from 3,412 patients against ABMIL, CLAM, ACMIL, TransMIL and MambaMIL, QLabelMIL reaches a macro-AUROC of up to 0.9186 with competitive calibration.

Figure 1. A frozen encoder produces patch embeddings used as keys and values. Label queries undergo self-attention (inter-label relationships) and cross-attention with the patches (per-label features Z). The head scores each label with (a) independent weights or (b) GCN-derived weights from the label co-occurrence graph.

---

Table of contents

• Method
• Results
• Repository structure
• Citation
• Acknowledgements
• License

---

Method

1. Shared query decoder

A frozen encoder gives patch features $H \in \mathbb{R}^{N \times d_e}$. They are projected and normalised into patch tokens, and $C$ learnable label queries $E$ are refined by $L$ decoder layers:

$$P = \mathrm{LN}(H W_p) \in \mathbb{R}^{N \times d}, \qquad Q^{(0)} = E \in \mathbb{R}^{C \times d}$$

$$Q' = Q^{(\ell-1)} + \mathrm{MHA}\big(\mathrm{LN},Q^{(\ell-1)}, \mathrm{LN},Q^{(\ell-1)}, \mathrm{LN},Q^{(\ell-1)}\big)$$

$$Q'' = Q' + \mathrm{MHA}\big(\mathrm{LN},Q', P, P\big)$$

$$Q^{(\ell)} = Q'' + \mathrm{FFN}\big(\mathrm{LN},Q''\big)$$

The output is $Z = \mathrm{LN}(Q^{(L)}) = [z_1, \dots, z_C]^\top$, where $z_c$ is the slide representation for pathology $c$.

• Self-attention is computed among the $C$ queries and captures inter-label relationships.
• Cross-attention has queries from the labels and keys/values from the patches. Its $C \times N$ attention matrix gives the per-class heatmaps, at cost $O(C \cdot N)$ rather than the $O(N^2)$ of patch-to-patch attention.

2. Classification head

QLabelMIL (independent). Each label has its own weight vector:

$$s_c = w_c^\top z_c + \beta_c, \qquad \hat p_c = \sigma(s_c)$$

QLabelMIL (GCN). The weights come from a GCN over a label co-occurrence graph built from the training labels:

$$M_{cc'} = \sum_n y_c^{(n)} y_{c'}^{(n)}, \qquad P_{cc'} = \frac{M_{cc'}}{M_{c'c'}} = \Pr(c \mid c')$$

An edge links $c$ and $c'$ when $P_{cc'} \ge \tau$. Each label keeps weight $1-p$ on itself and shares $p$ uniformly among its neighbours:

$$\hat A_{cc'} = \begin{cases} \dfrac{p}{\sum_{k \neq c} \mathbb{1}[P_{ck} \ge \tau]} & c \neq c',\ P_{cc'} \ge \tau \[2mm] 1-p & c = c' \ 0 & \text{otherwise} \end{cases}$$

$$G^{(\ell)} = \delta\big(\hat A, G^{(\ell-1)}, \Theta^{(\ell)}\big), \qquad [g_1, \dots, g_C]^\top = G^{(L_g)}, \qquad s_c = g_c^\top z_c + \beta_c$$

3. Loss

Both variants minimise binary cross-entropy averaged over labels, with a positive weight per class equal to the negative/positive ratio of the training split:

$$\mathcal{L} = \frac{1}{C} \sum_{c} \mathrm{BCE}(\hat p_c, y_c)$$

How QLabelMIL compares

Mechanism	ABMIL	CLAM	ACMIL	MambaMIL	TransMIL	QLabelMIL
Per-class attention maps	✗	✓	✗	✗	✗	✓
Class co-occurrence modelling	✗	✗	✗	✗	✗	✓
Label-relationship graph	✗	✗	✗	✗	✗	✓
Single multi-label model by design	✗	✗	✗	✗	✗	✓

_{Rows reflect the claims in the paper's Table 1 as discussed in its related-work section. Please verify against the final published table before release.}

---

Results

Dataset

5,764 H&E WSIs from 3,412 patients all scanned at the same facility. Slide-level labels were extracted from clinical reports. The split is by patient: 70% train, 15% validation, 15% test. Tissue was segmented with Otsu thresholding and tiled into non-overlapping 40× patches, with concentric 10× patches. Features were extracted at both magnifications with frozen ResNet18 and ResNet50 backbones.

Condition	Positives	Used for training
Intestinal metaplasia (IM)	1,372	✓
H. pylori (HP)	2,145	✓
Atrophy (AT)	1,838	✓
Chronic inflammation (CI)	4,882	✗ (no representative negatives)
Acute inflammation (AI)	32	✗ (too few positives)
Dysplasia (DY)	11	✗ (too few positives)

Models are trained on IM, HP and AT. Additional 6-class experiments gave similar outcomes and are not reported in the paper.

Diagnostic performance (test AUROC ↑)

R18 / R50 = ResNet18 / ResNet50 features. Bold = best, italic = second best in each column.

Pathology	Model	10× R18	10× R50	40× R18	40× R50
IM	ABMIL	0.9352	0.9356	0.9326	0.9433
	CLAM	0.9356	0.9377	0.9456	0.9528
	ACMIL	0.9269	0.9380	0.9489	0.9513
	MambaMIL	0.9327	0.9469	0.9453	0.9570
	TransMIL	0.9289	0.9377	0.9448	0.9495
	QLabelMIL	0.9427	0.9496	0.9456	0.9604
	QLabelMIL (GCN)	0.9342	0.9459	0.9430	0.9543
HP	ABMIL	0.8986	0.8969	0.9079	0.9051
	CLAM	0.9126	0.9149	0.9148	0.9120
	ACMIL	0.9088	0.9236	0.9018	0.9038
	MambaMIL	0.9008	0.9120	0.9034	0.9112
	TransMIL	0.9067	0.9179	0.9049	0.9114
	QLabelMIL	0.9082	0.9204	0.9163	0.9157
	QLabelMIL (GCN)	0.9147	0.9117	0.9194	0.9204
AT	ABMIL	0.8714	0.8654	0.8575	0.8660
	CLAM	0.8601	0.8618	0.8560	0.8677
	ACMIL	0.8498	0.8727	0.8600	0.8612
	MambaMIL	0.8563	0.8730	0.8453	0.8714
	TransMIL	0.8606	0.8827	0.8627	0.8735
	QLabelMIL	0.8643	0.8857	0.8632	0.8684
	QLabelMIL (GCN)	0.8642	0.8721	0.8595	0.8711
Macro avg.	ABMIL	0.9017	0.8993	0.8993	0.9048
	CLAM	0.9028	0.9048	0.9055	0.9108
	ACMIL	0.8952	0.9114	0.9036	0.9054
	MambaMIL	0.8966	0.9107	0.8980	0.9132
	TransMIL	0.8987	0.9128	0.9041	0.9115
	QLabelMIL	0.9051	0.9186	0.9083	0.9148
	QLabelMIL (GCN)	0.9044	0.9099	0.9073	0.9153

Takeaways

• A QLabelMIL variant is first in 12 of the 16 AUROC columns, including all four macro-average settings.
• Best macro-AUROC is 0.9186 (10×, ResNet50), +0.0058 over the strongest baseline (TransMIL).
• The independent variant has the best overall diagnostic performance. The GCN helps most for H. pylori, where it is first in three of four settings.
• IM is the easiest target and atrophy the hardest. ResNet50 generally beats ResNet18. Atrophy prefers 10×, IM and H. pylori prefer 40×.

Calibration (macro-ECE ↓)

Model	10× R18	10× R50	40× R18	40× R50
ABMIL	0.0917	0.1217	0.2170	0.0560
CLAM	0.1079	0.1052	0.0562	0.0497
ACMIL	0.0717	0.0784	0.1502	0.1089
MambaMIL	0.1203	0.1051	0.0604	0.0697
TransMIL	0.1617	0.0927	0.1016	0.1067
QLabelMIL	0.0575	0.0743	0.0781	0.0747
QLabelMIL (GCN)	0.1133	0.1195	0.0736	0.0882

QLabelMIL has the best macro-ECE in two of four settings and is competitive in the others. Per-pathology ECE is in Table 4 of the paper.

Training setup

AdamW with linear warm-up followed by cosine annealing; batch size 1 (variable patch count) with optional gradient accumulation; up to 200 epochs with early stopping (patience 20) on validation AUROC; independent hyper-parameter tuning per model; one NVIDIA RTX 3080.

Training

QLabelMIL (independent head and GCN head) can be plugged and ran from any pipeline, such as CLAM or ACMIL

Hyper-parameters reported in the paper as tunable: number of decoder layers $L$, embedding size $d$, number of GCN layers $L_g$ and its input size $d_0$, edge threshold $\tau$, neighbour weight $p$, learning rate, and gradient accumulation steps.

Repository structure

├── qlabelmil.py #architecture of the aggregator
├── utils.py 
├── architecture.png # figure used in this README
└── README.md

Citation

If you use this work, please cite:

@InProceedings{NetPed_QLabelMIL_MICCAISAT2026,
        author = { Neto, Pedro C. AND Lopes, Rita N. AND Prado e Castro, Lígia},
        title = { { QLabelMIL: Inter-Pathology Query Decoding for Multi-Label Gastric Histopathology } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17251},
        month = {pending},
        page = {pending}
}

Acknowledgements

QLabelMIL is inspired by Query2Label. Baselines compared in the paper: ABMIL, CLAM, ACMIL, TransMIL (NeurIPS 2021) and MambaMIL (MICCAI 2024).

Developed at Unilabs (Unilabs.AI) in collaboration with the Faculty of Engineering of the University of Porto (FEUP).

License

TODO: choose a license (e.g. MIT, Apache-2.0, or a non-commercial licence) and add a LICENSE file.