DeCo-MIL: Debiased Counterfactual Reasoning for Long-Tailed Whole Slide Image Analysis

Xiaoxiao Li; Xitong Ling; Jiawen Li; Weiming Chen; Zhenyang Cai; Xidong Wang; Tian Guan; Benyou Wang

cs.CVAug 12, 2026

DeCo-MIL: Debiased Counterfactual Reasoning for Long-Tailed Whole Slide Image Analysis

Authors: Xiaoxiao Li, Xitong Ling, Jiawen Li, Weiming Chen, Zhenyang Cai, Xidong Wang, Tian Guan, Benyou Wang, +1 more

Organizations: Tsinghua University · The Chinese University of Hong Kong, Shenzhen

Abstract

Multiple instance learning (MIL) is widely used for weakly supervised whole slide image (WSI) analysis. However, under long-tailed distributions, MIL-based WSI analysis faces a nested dual long-tail: an inter-slide class long tail and an intra-slide long tail of instance-level discriminative evidence. The two long tails are coupled: tail classes have few training slides, while their limited diagnostic evidence is concentrated in a few patches and obscured by abundant within-bag redundancy. This coupling biases models toward head classes and degrades rare-class recognition. To address this, we propose DeCo-MIL for long-tailed WSI analysis, which jointly alleviates the nested dual long-tail through frequency-debiased counterfactual reasoning. For the inner long tail, DeCo-MIL clusters patches into tissue-morphology anchors, replaces each anchor with its matched normal prototype to perform a counterfactual intervention, and estimates its counterfactual contribution to the ground-truth class using class-frequency-corrected predictions. These contributions guide redundancy masking to preserve scarce discriminative instances. For the outer long tail, DeCo-MIL constructs anchor-stratified pseudo-bags from redundancy-reduced bags and combines tail-aware oversampling with consistency regularization, increasing effective supervision for tail classes while preserving tissue-morphology composition. Extensive experiments on three long-tailed WSI benchmarks demonstrate that DeCo-MIL achieves state-of-the-art performance in both tail-class recognition and overall classification.

Figures & tables

Figure 1: The nested dual long-tail: (a) the inter-slide class long tail, (b) the intra-slide long tail of instance-level discriminative evidence.

Figure 2: Overview of the DeCo-MIL training framework. (a) WSI preprocessing and prototype clustering. (b) Debiased counterfactual scoring maps anchor contributions to masking ratios. (c) Tail-class oversampling and anchor-stratified sampling construct pseudo-bags. (d) The main and pseudo-bag auxiliary branches jointly train the final MIL model.

Method	Camelyon+-LT (IR=30)				PANDA-LT (IR=50)				TissueNet-LT (IR=10)
Method	ACC(%)	AUC(%)	F1(%)	Tail-F1(%)	ACC(%)	AUC(%)	F1(%)	Tail-F1(%)	ACC(%)	AUC(%)	F1(%)	Tail-F1(%)
AB-MIL	$61.56_{\scriptscriptstyle 2.54}$	$87.76_{\scriptscriptstyle 1.69}$	$52.46_{\scriptscriptstyle 2.71}$	$0.00_{\scriptscriptstyle 0.00}$	$51.28_{\scriptscriptstyle 0.83}$	$86.74_{\scriptscriptstyle 0.36}$	$45.90_{\scriptscriptstyle 1.30}$	$19.85_{\scriptscriptstyle 3.38}$	$56.67_{\scriptscriptstyle 1.05}$	$83.06_{\scriptscriptstyle 0.96}$	$49.64_{\scriptscriptstyle 0.81}$	$0.00_{\scriptscriptstyle 0.00}$
CLAM-SB	$63.75_{\scriptscriptstyle 1.53}$	$89.43_{\scriptscriptstyle 0.59}$	$54.54_{\scriptscriptstyle 1.61}$	$0.00_{\scriptscriptstyle 0.00}$	$49.48_{\scriptscriptstyle 0.50}$	$86.80_{\scriptscriptstyle 0.32}$	$41.74_{\scriptscriptstyle 0.84}$	$6.76_{\scriptscriptstyle 3.22}$	$55.50_{\scriptscriptstyle 0.41}$	$81.96_{\scriptscriptstyle 0.43}$	$48.42_{\scriptscriptstyle 0.30}$	$0.00_{\scriptscriptstyle 0.00}$
CLAM-MB	$61.88_{\scriptscriptstyle 0.77}$	$80.46_{\scriptscriptstyle 0.36}$	$52.84_{\scriptscriptstyle 0.82}$	$0.00_{\scriptscriptstyle 0.00}$	$51.40_{\scriptscriptstyle 0.84}$	$86.93_{\scriptscriptstyle 0.28}$	$45.70_{\scriptscriptstyle 1.68}$	$18.40_{\scriptscriptstyle 5.21}$	$59.50_{\scriptscriptstyle 0.85}$	$83.82_{\scriptscriptstyle 0.57}$	$52.75_{\scriptscriptstyle 1.47}$	$3.67_{\scriptscriptstyle 4.87}$
DS-MIL	$64.69_{\scriptscriptstyle 3.64}$	$88.65_{\scriptscriptstyle 2.31}$	$55.67_{\scriptscriptstyle 3.55}$	$0.00_{\scriptscriptstyle 0.00}$	$57.40_{\scriptscriptstyle 1.23}$	$88.32_{\scriptscriptstyle 0.35}$	$55.18_{\scriptscriptstyle 1.50}$	$41.68_{\scriptscriptstyle 6.15}$	$62.83_{\scriptscriptstyle 2.51}$	$86.16_{\scriptscriptstyle 1.86}$	$57.47_{\scriptscriptstyle 1.70}$	$15.48_{\scriptscriptstyle 4.37}$
TransMIL	$63.12_{\scriptscriptstyle 3.06}$	$83.82_{\scriptscriptstyle 2.48}$	$56.59_{\scriptscriptstyle 4.30}$	$14.37_{\scriptscriptstyle 9.74}$	$53.35_{\scriptscriptstyle 1.49}$	$85.46_{\scriptscriptstyle 1.10}$	$51.08_{\scriptscriptstyle 2.31}$	$40.61_{\scriptscriptstyle 8.21}$	$63.33_{\scriptscriptstyle 2.64}$	$86.36_{\scriptscriptstyle 1.61}$	$60.53_{\scriptscriptstyle 3.29}$	$26.96_{\scriptscriptstyle 10.54}$

Table 1: Comparison with baseline methods on three long-tailed pathology datasets. Subscripts denote standard deviations.

Figure 3: Overall and group-wise F1 scores of different methods on three long-tailed pathology datasets.

Dataset	Module	ACC(%)	AUC(%)	F1(%)	Tail-F1(%)
Camelyon+	Base	$61.56_{\scriptscriptstyle 2.54}$	$87.76_{\scriptscriptstyle 1.69}$	$52.46_{\scriptscriptstyle 2.71}$	$0.00_{\scriptscriptstyle 0.00}$
	+ Counterfactual	$65.94_{\scriptscriptstyle 1.17}$	$85.27_{\scriptscriptstyle 0.53}$	$59.45_{\scriptscriptstyle 1.60}$	$13.86_{\scriptscriptstyle 4.18}$
	+ Debiasing	$64.38_{\scriptscriptstyle 1.82}$	$89.67_{\scriptscriptstyle 0.86}$	$58.18_{\scriptscriptstyle 1.85}$	$18.04_{\scriptscriptstyle 5.12}$
	+ Pseudo-bag	$63.75_{\scriptscriptstyle 1.82}$	$90.18_{\scriptscriptstyle 1.19}$	$58.96_{\scriptscriptstyle 1.72}$	$29.16_{\scriptscriptstyle 4.10}$
	+ Oversampling	$\mathbf{68.44}_{\scriptscriptstyle\mathbf{1.17}}$	$\mathbf{91.08}_{\scriptscriptstyle\mathbf{0.36}}$	$\mathbf{64.65}_{\scriptscriptstyle\mathbf{1.11}}$	$\mathbf{36.63}_{\scriptscriptstyle\mathbf{4.13}}$
PANDA	Base	$51.28_{\scriptscriptstyle 0.83}$	$86.74_{\scriptscriptstyle 0.36}$	$45.90_{\scriptscriptstyle 1.30}$	$19.85_{\scriptscriptstyle 3.38}$

Table 2: Component ablation on three long-tailed pathology datasets.

Strategy	Camelyon+-LT		PANDA-LT		TissueNet-LT
Strategy	F1(%)	Tail-F1(%)	F1(%)	Tail-F1(%)	F1(%)	Tail-F1(%)
Random	$56.21_{\scriptscriptstyle 2.25}$	$25.09_{\scriptscriptstyle 7.54}$	$57.94_{\scriptscriptstyle 1.21}$	$48.37_{\scriptscriptstyle 1.82}$	$63.20_{\scriptscriptstyle 0.59}$	$38.57_{\scriptscriptstyle 2.18}$
Attention	$58.16_{\scriptscriptstyle 1.51}$	$20.34_{\scriptscriptstyle 7.52}$	$57.48_{\scriptscriptstyle 1.01}$	$46.81_{\scriptscriptstyle 1.43}$	$63.81_{\scriptscriptstyle 2.56}$	$41.91_{\scriptscriptstyle 3.28}$
Anchor-size	$61.30_{\scriptscriptstyle 3.46}$	$23.56_{\scriptscriptstyle 7.05}$	$58.83_{\scriptscriptstyle 0.71}$	$51.12_{\scriptscriptstyle 1.84}$	$65.46_{\scriptscriptstyle 1.89}$	$43.02_{\scriptscriptstyle 6.35}$
Counterfactual	$\mathbf{64.65}_{\scriptscriptstyle\mathbf{1.11}}$	$\mathbf{36.63}_{\scriptscriptstyle\mathbf{4.13}}$	$\mathbf{59.49}_{\scriptscriptstyle\mathbf{0.43}}$	$\mathbf{53.07}_{\scriptscriptstyle\mathbf{1.34}}$	$\mathbf{66.69}_{\scriptscriptstyle\mathbf{3.28}}$	$\mathbf{47.99}_{\scriptscriptstyle\mathbf{5.68}}$

Table 3: Comparison of different masking strategies on three long-tailed pathology datasets.

Strategy	Camelyon+-LT		PANDA-LT		TissueNet-LT
Strategy	F1(%)	Tail-F1(%)	F1(%)	Tail-F1(%)	F1(%)	Tail-F1(%)
Zero	$56.16_{\scriptscriptstyle 2.60}$	$21.30_{\scriptscriptstyle 13.13}$	$57.61_{\scriptscriptstyle 0.92}$	$48.26_{\scriptscriptstyle 3.24}$	$64.21_{\scriptscriptstyle 2.11}$	$41.01_{\scriptscriptstyle 4.50}$
Anchor Proto.	$61.22_{\scriptscriptstyle 3.41}$	$25.29_{\scriptscriptstyle 9.37}$	$58.28_{\scriptscriptstyle 1.11}$	$51.90_{\scriptscriptstyle 1.20}$	$64.53_{\scriptscriptstyle 2.64}$	$43.93_{\scriptscriptstyle 4.84}$
Normal Proto.	$\mathbf{64.65}_{\scriptscriptstyle\mathbf{1.11}}$	$\mathbf{36.63}_{\scriptscriptstyle\mathbf{4.13}}$	$\mathbf{59.49}_{\scriptscriptstyle\mathbf{0.43}}$	$\mathbf{53.07}_{\scriptscriptstyle\mathbf{1.34}}$	$\mathbf{66.69}_{\scriptscriptstyle\mathbf{3.28}}$	$\mathbf{47.99}_{\scriptscriptstyle\mathbf{5.68}}$

Table 4: Comparison of different counterfactual constructions on three long-tailed pathology datasets.

Strategy	Camelyon+-LT		PANDA-LT		TissueNet-LT
Strategy	F1(%)	Tail-F1(%)	F1(%)	Tail-F1(%)	F1(%)	Tail-F1(%)
Random	$58.20_{\scriptscriptstyle 3.44}$	$13.11_{\scriptscriptstyle 12.41}$	$57.19_{\scriptscriptstyle 0.54}$	$47.51_{\scriptscriptstyle 1.09}$	$64.35_{\scriptscriptstyle 1.82}$	$38.48_{\scriptscriptstyle 4.49}$
Equal-per-Anchor	$61.31_{\scriptscriptstyle 4.27}$	$22.72_{\scriptscriptstyle 14.89}$	$58.17_{\scriptscriptstyle 0.48}$	$50.47_{\scriptscriptstyle 1.56}$	$64.26_{\scriptscriptstyle 2.43}$	$42.17_{\scriptscriptstyle 3.20}$
Stratified	$\mathbf{64.65}_{\scriptscriptstyle\mathbf{1.11}}$	$\mathbf{36.63}_{\scriptscriptstyle\mathbf{4.13}}$	$\mathbf{59.49}_{\scriptscriptstyle\mathbf{0.43}}$	$\mathbf{53.07}_{\scriptscriptstyle\mathbf{1.34}}$	$\mathbf{66.69}_{\scriptscriptstyle\mathbf{3.28}}$	$\mathbf{47.99}_{\scriptscriptstyle\mathbf{5.68}}$

Table 5: Effectiveness of different sampling strategies on three long-tailed pathology datasets.

Figure 4: Performance comparison under different imbalance ratios on PANDA-LT. (a) Total F1 and (b) Tail-F1 of AB-MIL and DeCo-MIL. The purple annotations denote the Tail-F1 percentage-point gains of DeCo-MIL over AB-MIL.

Figure 5: Nested dual long-tail analysis on PANDA-LT (

IR=50

): (a) ranked debiased counterfactual contributions of tissue anchors and (b) the corresponding Lorenz curves and Gini coefficients.

Figure 6: Visualization of DeCo-MIL on tail-class slides from Camelyon+-LT: From left to right: the original WSI with official lesion annotations, morphology-anchor assignments, debiased counterfactual contribution heatmaps, and expected retention maps. Yellow contours denote the official lesion annotations.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Head	Medium	Tail
Camelyon+-LT (IR=30)	0, 2	1	3
PANDA-LT (IR=50)	0, 1	2, 4	3, 5
TissueNet-LT (IR=10)	1, 0	2	3

Appendix

Table A.1: Class assignments to frequency groups.

Split	0	1	2	3	Total
Full dataset	870	174	251	54	1,349
Train (IR=30)	844	148	225	28	1,245
Validation	10	10	10	10	40
Test	16	16	16	16	64

Appendix

Table A.2: Camelyon+-LT data statistics.

Split	0	1	2	3	4	5	Total
Full dataset	2,603	2,399	1,209	1,118	1,124	1,102	9,555
Train (IR=50)	2,303	1,053	481	100	220	46	4,203
Validation	100	100	100	100	100	100	600
Test	200	200	200	200	200	200	1,200

Appendix

Table A.3: PANDA-LT data statistics under the main IR=50 setting.

Nominal IR	Label 0	Label 1	Label 2	Label 3	Label 4	Label 5	Train total	Realized IR
10	2,303	1,453	909	365	578	230	5,838	10.0
20	2,303	1,264	694	209	381	115	4,966	20.0
30	2,303	1,166	590	151	299	76	4,585	30.3
40	2,303	1,101	526	120	251	57	4,358	40.4
50	2,303	1,053	481	100	220	46	4,203	50.1

Appendix

Table A.4: PANDA-LT training-set statistics at different nominal imbalance ratios.

Split	0	1	2	3	Total
Full dataset	268	288	238	219	1,013
Train (IR=10)	110	238	51	23	422
Validation	20	20	20	20	80
Test	30	30	30	30	120

Appendix

Table A.5: TissueNet-LT data statistics under the main IR=10 setting.

Dataset	Label-0 category	Slides
Camelyon+-LT	Negative	844
PANDA-LT	Benign	2,303
TissueNet-LT	Normal/subnormal	110

Appendix

Table A.6: Training slides used to construct normal prototypes.

Construction	Replacement $b_{k}$
Zero replacement	$\mathbf{0}\in\mathbb{R}^{512}$
Morphology-prototype replacement	$H_{k}$
Matched-normal replacement (ours)	$\hat{q}_{k}$

Appendix

Table C.1: Definitions of the compared counterfactual constructions.

Figure D.1: Sensitivity to $\lambda$ . Each bar shows the mean over five runs.

Figure D.2: Sensitivity to the maximum masking ratio $r_{\max}$ . Each bar shows the mean over five runs.

Figure D.3: Sensitivity to the oversampling strength $\alpha$ . Each bar shows the mean over five runs.

Figure D.4: Additional visualizations on Camelyon+-LT. Each row shows a different correctly classified slide. From left to right: the original tissue image, morphology-anchor assignment, debiased counterfactual contribution, and expected retention map. Yellow contours denote official tumor annotations. Warm and cool colors indicate positive and negative contributions, respectively, while darker green indicates higher expected retention.

Explore similar work

One Tile, Multiple Instances: Rethinking MIL for Sparse Diagnostic Evidence
Oct 4, 2026Runsheng Liu, Cheng Jin, Hao Jiang +1Multiple Instance Learning Whole-Slide Image Analysis
Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong, China. · Department of Chemical and Biological Engineering, The Hong Kong University of Science and Technology, Hong Kong, China. · Division of Life Science, The Hong Kong University of Science and Technology, Hong Kong, China. +3
GCE-MIL: Faithful and Recoverable Evidence for Multiple Instance Learning in Whole-Slide Imaging
May 17, 2026Xiangyu Li, Ran SuExplainable Medical Image Analysis Histopathology Image Classification
College of Intelligence and Computing Tianjin University
LRMIL: Efficient Low-Resolution Multiple Instance Learning via High-Resolution Knowledge Distillation for Whole Slide Image Classification
Jun 5, 2026Yonghan Shin, Won-Ki JeongHistopathology Image Classification Knowledge Distillation
Department of Computer Science and Engineering, Korea University, Seoul, Korea

Oct 4, 2026J/K move · Enter open · S save

One Tile, Multiple Instances: Rethinking MIL for Sparse Diagnostic Evidence

Runsheng Liu, Cheng Jin, Hao Jiang, Hao Chen

Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong, China. · Department of Chemical and Biological Engineering, The Hong Kong University of Science and Technology, Hong Kong, China. · Division of Life Science, The Hong Kong University of Science and Technology, Hong Kong, China.+3

Multiple Instance Learning Whole-Slide Image Analysis Histopathology Image Classification

Open paper AskPDF arXiv

In weakly supervised Whole Slide Image (WSI) classification, feature extractors typically compress each image tile into a single global embedding. Consequently, slide-level aggregators are restricted to this coarse tile scale, concealing fine-grained sub-tile evidence from the attention mechanism. We introduce DI-MIL, a framework that decouples encoding context from instance granularity through decomposed instances. By clustering dense spatial tokens from a frozen foundation model within each tile, DI-MIL converts a single tile into multiple independently weighted instance embeddings. As a training-free post-encoding module, DI-MIL integrates seamlessly into existing pipelines without requiring re-encoding or downstream architectural modifications. We evaluate DI-MIL on cytopathology, a challenging testbed where sparse diagnostic signals are easily diluted within standard tiles. Across four datasets, three frozen foundation models, and two attention-based aggregators, DI-MIL demonstrates consistent efficacy, improving 67 of 72 metric-level comparisons, with the largest mean gains reaching 3.64 points under cytopathology-specific backbones. In a broader comparison against seven representative MIL baselines, DI-MIL paired with ACMIL achieves highest mean performance in 33 of 36 backbone-dataset-metric comparisons. Ablations show that direct smaller tiling inflates the extracted tile count by up to 43.3

\times

with non-monotonic performance, whereas DI-MIL incurs zero additional image-extraction overhead while achieving the strongest overall results. These results establish instance construction as an orthogonal design dimension in MIL, supporting DI-MIL as a cost-efficient solution under sparse diagnostic evidence.

Abstract

Explore similar work

One Tile, Multiple Instances: Rethinking MIL for Sparse Diagnostic Evidence

GCE-MIL: Faithful and Recoverable Evidence for Multiple Instance Learning in Whole-Slide Imaging

LRMIL: Efficient Low-Resolution Multiple Instance Learning via High-Resolution Knowledge Distillation for Whole Slide Image Classification