BagShift: Measuring How Patch Selection Changes the Evidence Seen by Whole-Slide MIL

Ruicheng Yuan; Zhenxuan Zhang; Liwei Hu; Anbang Wang; Haijie Xu; Jiawei Luo; Guang Yang

cs.CVAug 16, 2026

BagShift: Measuring How Patch Selection Changes the Evidence Seen by Whole-Slide MIL

Authors: Ruicheng Yuan, Zhenxuan Zhang, Liwei Hu, Anbang Wang, Haijie Xu, Jiawei Luo, Guang Yang

Organizations: College of Computer Science and Electronic Engineering, Hunan University, Changsha, Hunan, China · Department of Bioengineering and Imperial-X, Imperial College London, London, UK

Abstract

Whole-slide multiple-instance learning (MIL) observes only the patches admitted by its selector. Deployment can alter this selector through compute limits, tissue masking, or regional workflows, even when the patch count is unchanged. We introduce BagShift, a paired protocol that changes the selector for the same case while holding its features and predictor fixed, thereby isolating selector response from case mix. With equal 128-patch budgets, sampling across the tissue or concentrating around one coordinate exposes markedly different evidence: on PANDA, the two views reduce quadratic weighted kappa by 1.57 and 17.96 points, respectively (QWK reported on the $\times100$ scale). On CAMELYON16, lesion annotations withheld from model development show that localized views retain tumor in only 10.0% of micrometastatic observations, and matched exposure does not consistently recover the loss. The same fixed-count stressor produces a much smaller response on external lung subtyping, although differences in relative coverage make cross-task severity descriptive. When repeated localized observations are available, unioning their patches before one nonlinear MIL pass improves PANDA QWK by 7.87 points over averaging regional predictions. Patch count specifies computation, not observed evidence; deployment evaluations should report both what a selector preserves and how repeated observations are aggregated.

Figures & tables

Figure 1: Equal patch budgets can retain different focal evidence. (a) The same CAMELYON16 micrometastatic slide under reference, dispersed-128, and localized-128 selection; magenta marks annotated tumor and colored overlays mark retained patches. Both restricted bags contain 128 patches but retain one versus zero tumor-overlap patches. (b) Metastasis probability across five model seeds. (c) Tumor-retention frequency across 27 micrometastatic slides and ten selector draws. The deterministic case-selection rule is given in the supplement.

Figure 2: BagShift measures selector response by paired intervention. (a) Exact PANDA coordinates for the reference pool and equally sized dispersed and localized bags; the footprints depict sampling geometry rather than histology. (b) The intervention changes only the selector while holding the slide, label, features, and predictor fixed. (c) Five-seed mean $\Delta$ QWK for reference-only and validation-selected matched exposure ( $\alpha=1$ ); exposure reduces both responses but leaves a large geometry gap.

		Absolute QWK $\uparrow$			Degradation (Ref. $-$ policy) $\downarrow$
Control family	Training control	Ref.	Budget-128	Spatial-128	$\Delta$ Budget	$\Delta$ Spatial
Baseline	Reference only	94.13 $\pm$ 0.34	92.56 $\pm$ 0.81	76.17 $\pm$ 2.29	1.57 $\pm$ 0.78	17.96 $\pm$ 2.38
Deletion	Random deletion only	94.26 $\pm$ 0.27	93.17 $\pm$ 0.92	77.23 $\pm$ 1.47	1.09 $\pm$ 1.03	17.03 $\pm$ 1.57
Exposure	Spatial-policy only	93.96 $\pm$ 0.36	92.88 $\pm$ 0.96	79.30 $\pm$ 1.04	1.08 $\pm$ 0.73	14.66 $\pm$ 1.21
Exposure/weight	Matched average, $\alpha=0.5$	94.27 $\pm$ 0.24	93.33 $\pm$ 0.68	77.82 $\pm$ 1.52	0.93 $\pm$ 0.75	16.45 $\pm$ 1.75
Exposure/weight	Matched average, $\alpha=1$ †	94.29 $\pm$ 0.16	93.48 $\pm$ 0.91	78.98 $\pm$ 1.44	0.81 $\pm$ 0.79	15.30 $\pm$ 1.54

Table 1: PANDA test QWK on the

\times 100

scale (mean

\pm

SD across five seeds). The controls separate deletion, exposure, weight on restricted views, consistency, and reduction within each case. Higher absolute QWK and lower degradation are better;

\dagger

marks the operating point selected on validation data.

	AUROC				Balanced accuracy
Training	Ref.	Budget	Spatial	Sp. $-$ Bu.	Ref.	Budget	Spatial	Sp. $-$ Bu.
Reference only	99.19 $\pm$ 0.85	88.22 $\pm$ 1.19	68.50 $\pm$ 1.92	$-$ 19.72 $\pm$ 0.89	90.00 $\pm$ 5.12	80.21 $\pm$ 0.62	61.63 $\pm$ 0.11	$-$ 18.58 $\pm$ 0.60
Matched average	99.16 $\pm$ 0.55	88.08 $\pm$ 1.10	67.38 $\pm$ 2.02	$-$ 20.70 $\pm$ 1.61	92.24 $\pm$ 5.52	80.47 $\pm$ 1.01	61.71 $\pm$ 0.20	$-$ 18.76 $\pm$ 0.85

Table 2: CAMELYON16 response on the official test set (%, mean

\pm

SD across five seeds after averaging metrics for each draw within each seed). Budget and spatial views both contain 128 patches. Localized selection loses substantially more performance, and matched exposure does not recover it.

Training control	Ref.	Spatial	$\Delta$
Average, $\alpha=0.5$	96.74 $\pm$ 0.61	96.18 $\pm$ 0.64	0.56 $\pm$ 0.24
Average, $\alpha=1$	96.71 $\pm$ 0.44	96.08 $\pm$ 0.51	0.63 $\pm$ 0.39
Worst policy	97.04 $\pm$ 0.25	96.50 $\pm$ 0.29	0.54 $\pm$ 0.19

Table 3: CPTAC AUROC after patient aggregation (%, mean

\pm

SD across five seeds) after TCGA training.

\Delta

is reference minus spatial AUROC.

Figure 3: Aggregation order changes the response to the same acquired views. (a) Prediction averaging uses

K

independent MIL passes; union-before-MIL deduplicates patches and uses one shared pass. (b) PANDA

\Delta

QWK versus

K

(mean

\pm

SD across three seeds); the upper axis gives union coverage. At

K=5

, union-before-MIL gains 7.87 QWK points (paired 95% CI 6.07–9.82).

Figure S1: Detailed selector diagnostics. (a) Three-seed mean expected-grade shifts for budget and spatial selectors on PANDA. (b) CPTAC spatial-window true-class margin shifts with conditional patient-bootstrap intervals. Negative values denote movement away from the true class.

Figure S2: Two deterministically selected PANDA cases on level-1 H&E. The reference column overlays the retrospective Radboud grade mask; the middle columns show the exact 128-patch views. The tall first slide is folded into upper and lower halves without rescaling. All panels preserve tissue aspect ratio; white padding only standardizes the display cells. Masks were unavailable to model and selector development; predictions use the unchanged validation-selected Attention control. In the affected case, reference and Budget-128 distributions nearly coincide while Spatial-128 shifts to grade 3; in the stable case, all three distributions nearly coincide by the frozen case-selection rule.

Model	Reference	Spatial 128	$\Delta$
CLAM-SB	96.40 $\pm$ 1.36	95.51 $\pm$ 1.93	0.90 $\pm$ 0.58
CLAM-SB + views	97.13 $\pm$ 0.07	96.56 $\pm$ 0.36	0.56 $\pm$ 0.29
DSMIL	96.48 $\pm$ 0.23	95.40 $\pm$ 0.08	1.08 $\pm$ 0.16
DSMIL + views	96.26 $\pm$ 0.44	95.36 $\pm$ 0.48	0.91 $\pm$ 0.28
TransMIL	96.85 $\pm$ 0.49	96.11 $\pm$ 0.50	0.73 $\pm$ 0.21
TransMIL + views	95.18 $\pm$ 0.27	94.18 $\pm$ 0.45	1.00 $\pm$ 0.18

Table S1: CPTAC patient AUROC after TCGA training (%).

Model	Objective	Reference	Spatial	$\Delta$
CLAM-SB	Average	93.44 $\pm$ 0.60	77.52 $\pm$ 1.84	15.92 $\pm$ 1.63
CLAM-SB	Worst	93.80 $\pm$ 0.53	78.72 $\pm$ 1.40	15.08 $\pm$ 1.40
DSMIL	Average	94.44 $\pm$ 0.61	77.36 $\pm$ 2.04	17.08 $\pm$ 1.96
DSMIL	Worst	94.03 $\pm$ 0.52	77.97 $\pm$ 0.96	16.05 $\pm$ 1.05
TransMIL	Average	94.73 $\pm$ 0.27	79.15 $\pm$ 1.64	15.58 $\pm$ 1.44
TransMIL	Worst	94.54 $\pm$ 0.58	79.23 $\pm$ 1.57	15.31 $\pm$ 1.49

Table S2: Five-seed PANDA QWK on the

\times 100

scale for matched-average

\alpha=0.5

and worst-policy objectives using identical views.

Control	Ref. AUC/BAcc	Spatial AUC/BAcc	$\Delta$
Average, $\alpha=0.5$	96.74 $\pm$ 0.61/90.49 $\pm$ 1.52	96.18 $\pm$ 0.64/90.13 $\pm$ 1.61	0.56 $\pm$ 0.24
Average, $\alpha=1$	96.71 $\pm$ 0.44/89.67 $\pm$ 1.34	96.08 $\pm$ 0.51/89.40 $\pm$ 1.51	0.63 $\pm$ 0.39
Worst policy	97.04 $\pm$ 0.25/90.57 $\pm$ 1.28	96.50 $\pm$ 0.29/89.66 $\pm$ 1.52	0.54 $\pm$ 0.19

Table S3: Untouched CPTAC patient results after TCGA training (%, five-seed mean

\pm

SD). Cells report AUROC/BAcc;

\Delta

is reference minus spatial AUROC.

Variant	PANDA	CPTAC ref.	CPTAC spatial
Attention	74.88	96.73 / 89.73	95.72 / 88.53
+ policy views	77.20	96.82 / 90.08	96.00 / 89.31
+ CVaR	79.73	96.83 / 90.13	96.13 / 89.37
+ EMA loss	79.68	97.10 / 90.96	96.43 / 90.25

Table S4: Earlier Attention component means on

\times 100

scales. PANDA reports spatial QWK; CPTAC reports patient AUROC / balanced accuracy.

Figure S3: Fixed-weight sensitivity. (a) One global

\alpha

is selected by mean QWK over five validation policies, five model seeds, and ten selector draws. (b) Spatial-128 test response; the dashed line marks worst-policy performance. (c) Absolute reference QWK across the same weights. Error bars are SD across seeds; the shaded band and dotted line mark the selected

\alpha

Control	Val. mean	Reference	Spatial	$\Delta$
Average, $\alpha=0.25$	91.09 $\pm$ 0.24	94.27 $\pm$ 0.29	77.44 $\pm$ 1.76	16.83 $\pm$ 1.99
Average, $\alpha=0.50$	91.23 $\pm$ 0.20	94.27 $\pm$ 0.24	77.82 $\pm$ 1.52	16.45 $\pm$ 1.75
Average, $\alpha=1.00$	91.36 $\pm$ 0.12	94.29 $\pm$ 0.16	78.98 $\pm$ 1.44	15.30 $\pm$ 1.54
Average, $\alpha=1.50$	91.32 $\pm$ 0.18	94.22 $\pm$ 0.35	79.26 $\pm$ 1.37	14.96 $\pm$ 1.69
Average, $\alpha=2.00$	91.33 $\pm$ 0.19	94.21 $\pm$ 0.31	79.51 $\pm$ 1.20	14.70 $\pm$ 1.42
Worst policy	–	94.39 $\pm$ 0.37	79.65 $\pm$ 1.15	14.74 $\pm$ 1.37

Table S5: Complete fixed-weight QWK curve on the

\times 100

scale (five-seed mean

\pm

SD). Validation is the mean QWK over five policies; remaining columns are PANDA test results.

Control	Ref.	Sp. $\Delta$	Draw SD	Max $\Delta$	Stripe $\Delta$
Spatial-policy only	93.96	14.01	1.41	16.13	11.88
Validation-selected $\alpha=1$	94.29	14.88	1.36	17.01	12.58
Sensitivity endpoint $\alpha=2$	94.21	14.22	1.44	16.36	12.05
Policy-wise worst	94.39	14.03	1.33	16.12	11.95

Table S6: PANDA selector variability over five model seeds and ten draws per seed. Ref. is reference QWK; remaining policy values are QWK degradation points, and SD is across selector draws.

$K$	Coverage	Localized prediction average	Localized patch union	Dispersed prediction average	Coverage-matched random
1	32.8%	$13.47\pm 1.38$	$13.47\pm 1.38$	$1.31\pm 0.94$	$1.43\pm 0.31$
2	51.4%	$12.72\pm 1.46$	$5.50\pm 0.98$	$0.22\pm 0.23$	$1.02\pm 0.44$
5	79.9%	$9.16\pm 0.73$	$1.29\pm 0.56$	$0.04\pm 0.51$	$0.42\pm 0.51$
10	94.4%	$9.43\pm 0.25$	$0.18\pm 0.16$	$-0.18\pm 0.15$	$0.12\pm 0.27$
20	99.3%	$8.68\pm 0.95$	$0.08\pm 0.12$	$-0.08\pm 0.10$	$0.01\pm 0.03$

Table S7: PANDA QWK degradation in points (three-seed mean

\pm

SD) as the number of localized observations increases. Lower is better; negative values indicate a small improvement over the reference view. At

K=5

, localized union improves over localized prediction averaging by 7.87 points (paired 95% CI 6.07–9.82).

Task	Strategy	Skill loss	Directional shift
PANDA	Spatial average	16.45%	$-0.0408$
PANDA	Spatial union	0.08%	$+0.0000$
CPTAC	Spatial average	0.99%	$-0.0097$
CPTAC	Spatial union	0.00%	$-0.0000$

Table S8: Twenty-draw mechanism check. Skill loss and directional shift compare the restricted strategy with its reference prediction.

Training	Policy	AUROC	AUPRC	Balanced accuracy
Reference only	Full	$0.9919\pm 0.0085$	$0.9898\pm 0.0092$	$0.9000\pm 0.0512$
	Random keep 75%	$0.9902\pm 0.0095$	$0.9884\pm 0.0099$	$0.8980\pm 0.0404$
	Random keep 50%	$0.9844\pm 0.0084$	$0.9832\pm 0.0081$	$0.8943\pm 0.0352$
	Random budget 128	$0.8822\pm 0.0119$	$0.8828\pm 0.0079$	$0.8021\pm 0.0062$
	Spatial window 128	$0.6850\pm 0.0192$	$0.6647\pm 0.0150$	$0.6163\pm 0.0011$
Matched average	Full	$0.9916\pm 0.0055$	$0.9894\pm 0.0068$	$0.9224\pm 0.0552$

Table S9: CAMELYON16 official-test policy response. Metrics are proportions; budget and spatial views contain 128 patches.

Training	View	Stage	Sensitivity	AUROC vs. normal	Mean positive probability
Reference only	Full	Macro	$1.000\pm 0.000$	$1.000\pm 0.000$	$0.729$
	Full	Micro	$0.637\pm 0.168$	$0.985\pm 0.014$	$0.568$
	Budget 128	Macro	$0.982\pm 0.030$	$0.994\pm 0.011$	$0.719$
	Budget 128	Micro	$0.299\pm 0.063$	$0.791\pm 0.034$	$0.406$
	Spatial 128	Macro	$0.435\pm 0.062$	$0.747\pm 0.062$	$0.471$
	Spatial 128	Micro	$0.073\pm 0.058$	$0.634\pm 0.054$	$0.308$

Table S10: CAMELYON16 stage response (mean

\pm

SD across 50 nested seed/draw cells). The full view is deterministic within a model seed.

$K$	Median retained	Component AUROC	Matched-random AUROC	Difference	Micro tumor retained component / random
1	44.9%	$92.02\pm 1.46$	$97.64\pm 0.84$	$-5.62\pm 1.12$	74.1% / 99.6%
2	79.8%	$99.17\pm 0.61$	$98.82\pm 0.82$	$+0.35\pm 0.31$	88.9% / 99.6%
3	89.3%	$99.14\pm 0.68$	$99.04\pm 0.95$	$+0.10\pm 0.34$	96.3% / 100.0%
5	96.2%	$99.28\pm 0.62$	$99.11\pm 0.93$	$+0.17\pm 0.32$	100.0% / 100.0%
10	98.4%	$99.22\pm 0.78$	$99.18\pm 0.86$	$+0.03\pm 0.08$	100.0% / 100.0%

Table S11: Exploratory CAMELYON16 tissue-component cleanup. AUROC is in percentage points (five-seed mean

\pm

SD); random metrics first average ten draws within each seed. Difference is component minus exact-cardinality random. At

K=1

, all five seed differences are negative. A label-stratified paired bootstrap on seed/draw-averaged slide predictions gives

-7.55

AUROC points (95% CI

-14.18

-2.17

$K$	Mean patches	Coverage	Prediction average	Union before MIL	Cardinality-matched random
1	128	3.45%	$0.6202\pm 0.0319$ / $0.5857\pm 0.0056$	$0.6202\pm 0.0319$ / $0.5857\pm 0.0056$	$0.8831\pm 0.0114$ / $0.8060\pm 0.0089$
2	250	6.60%	$0.6971\pm 0.0396$ / $0.6122\pm 0.0414$	$0.7161\pm 0.0301$ / $0.6579\pm 0.0051$	$0.8865\pm 0.0070$ / $0.8134\pm 0.0211$
5	601	15.62%	$0.8139\pm 0.0162$ / $0.5857\pm 0.0056$	$0.8282\pm 0.0143$ / $0.7735\pm 0.0133$	$0.9133\pm 0.0129$ / $0.8531\pm 0.0256$
10	1,105	27.45%	$0.8327\pm 0.0180$ / $0.5980\pm 0.0091$	$0.8443\pm 0.0105$ / $0.8000\pm 0.0056$	$0.9660\pm 0.0075$ / $0.8988\pm 0.0244$
20	1,900	44.41%	$0.8968\pm 0.0139$ / $0.6041\pm 0.0046$	$0.9116\pm 0.0148$ / $0.8571\pm 0.0250$	$0.9701\pm 0.0085$ / $0.9041\pm 0.0406$

Table S12: CAMELYON16 repeated-window response under five matched-model seeds. Entries are AUROC / balanced accuracy; coverage is the fraction of each reference bag represented by the exact union.

Task ( $K=5$ )	Aggregation	Spearman $\uparrow$	Macro AUROC $\uparrow$	Log loss $\downarrow$	RPS $\downarrow$	Sensitivity $\uparrow$
PANDA	Prediction average	$0.9019\pm 0.0008$	$93.62\pm 0.24$	$0.810\pm 0.003$	$0.0617\pm 0.0002$	–
PANDA	Union before MIL	$0.9331\pm 0.0025$	$95.45\pm 0.26$	$0.635\pm 0.016$	$0.0432\pm 0.0014$	–
CAMELYON16	Prediction average	–	$81.39\pm 1.62$	$0.581\pm 0.005$	–	$17.1\pm 1.1$
CAMELYON16	Union before MIL	–	$82.82\pm 1.43$	$0.482\pm 0.020$	–	$54.7\pm 2.7$

Table S13: Probability-level aggregation diagnostics (model-seed mean

\pm

SD). PANDA macro AUROC is one-versus-rest and RPS is the normalized ranked probability score. CAMELYON16 specificity is 100% for both aggregation orders at the fixed 0.5 threshold.

			True-class probability			Balanced accuracy
Hospital	All cases	Paired cases	Evidence+	Random	Evidence $-$	Evidence+	Random	Evidence $-$
H1	1,981	202	46.92 $\pm$ 1.49	38.28 $\pm$ 1.10	30.94 $\pm$ 0.70	59.09 $\pm$ 1.18	48.77 $\pm$ 1.73	40.63 $\pm$ 2.52
H2	3,981	242	53.53 $\pm$ 0.99	42.07 $\pm$ 0.51	30.81 $\pm$ 0.55	71.95 $\pm$ 2.29	57.65 $\pm$ 1.36	41.90 $\pm$ 2.02

Table S14: Two-hospital human-evidence selection (%, five-seed mean

\pm

SD; ten selector draws). Positive and negative views each contain one image from the same held-out case.

Hospital	Cases	Median	Mean	Max.
H1	5,000	2	2.088	27
H2	5,000	2	2.010	6
H3	5,000	2	2.258	29

Table S15: Frozen private cohort.

Hospital	Policy	I2T R@1	Drop	Matched cosine	Drop
H1	Full	12.76 $\pm$ 0.89	–	50.22 $\pm$ 1.10	–
	Random single	11.07 $\pm$ 1.09	1.69	47.66 $\pm$ 1.13	2.56
	Keep 50%	11.24 $\pm$ 1.04	1.52	47.93 $\pm$ 1.12	2.29
	Top-attention drop 50%	10.76 $\pm$ 1.37	2.00	47.22 $\pm$ 1.15	3.00
H2	Full	10.69 $\pm$ 1.38	–	54.98 $\pm$ 1.08	–
	Random single	7.83 $\pm$ 1.54	2.86	49.73 $\pm$ 1.08	5.25

Table S16: Attention MIL image-removal response (%). Drops are relative to the full multi-image case; larger positive drops indicate greater sensitivity.

Hospital	Policy	Attention	Earlier retrieval objective	Difference	Favorable seeds
H1	Full	12.76 $\pm$ 0.89	12.09 $\pm$ 1.08	$-0.67$	0/3
	Random single	11.07 $\pm$ 1.09	10.82 $\pm$ 0.91	$-0.25$	2/3
	Keep 50%	11.24 $\pm$ 1.04	10.92 $\pm$ 0.86	$-0.32$	2/3
	Top-attention drop 50%	10.76 $\pm$ 1.37	10.36 $\pm$ 0.28	$-0.40$	2/3
H2	Full	10.69 $\pm$ 1.38	11.97 $\pm$ 0.41	$+1.28$	3/3
	Random single	7.83 $\pm$ 1.54	8.25 $\pm$ 0.43	$+0.41$	2/3

Table S17: Equivalence-aware image-to-text Recall@1 (%). Positive differences favor the earlier retrieval objective, which is distinct from the BagShift measurement framework studied in the main paper.

Whole-Slide Image Analysis

Whole-slide image analysis uses AI to interpret gigapixel pathology slides for diagnosis, prognosis, tissue mapping, and report generation.

See all 214 papers

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
Spatial Blindness in Whole-Slide Multiple Instance Learning
May 17, 2026Xiangyu Li, Ran SuHistopathology Image Classification Multiple Instance Learning
College of Intelligence and Computing Tianjin University Tianjin, China
Mixture of Mini Experts: Overcoming the Linear Layer Bottleneck in Multiple Instance Learning
Mar 23, 2026Daniel Shao, Joel Runevic, Richard J. Chen +4Multiple Instance Learning Histopathology Image Classification
Massachusetts Institute of Technology, Cambridge MA · Harvard University, Cambridge MA · Emory University, Atlanta GA +1

May 17, 2026J/K move · Enter open · S save

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

Xiangyu Li, Ran Su

College of Intelligence and Computing Tianjin University

Explainable Medical Image Analysis Histopathology Image Classification Multiple Instance Learning Whole-Slide Image Analysis

Open paper AskPDF arXiv

Multiple instance learning (MIL) is the standard approach for whole-slide image (WSI) classification and survival prediction, where attention-based models ag gregate patch features into slide-level predictions. These models treat attention weights as evidence for their predictions, but attention is optimized for classi fication, not for identifying which patches actually support the diagnosis. This conflation leads to three failures: selected patches are insufficient (keeping them alone drops Macro-F1 by 0.078), unnecessary (removing them barely changes the prediction), and unrecoverable (continuous attention scores disagree with discrete patch subsets used at inference). The central premise is that evidence quality should be optimized directly through explicit criteria- Sufficiency, Necessity, and Recov erability (S/N/R)- rather than inherited as a byproduct of classification. GCE-MIL is a backbone-agnostic wrapper implemented through three injection modes and three evidence components: a grounding mechanism that aligns selection with domain-specific concepts, noisy-OR coverage that acts as a differentiable proxy for interventional evidence search, and threshold-plus-repair recovery that converts continuous selectors into discrete subsets through marginal-guided repair. Across 9 backbones and 9 datasets (81 configurations), GCE-MIL improves average Macro-F1 by 0.024 and C-index by 0.014, reduces the continuous-discrete gap by 4-7, and increases complement degradation by 2-4. With optional tile prefiltering after discrete recovery, inference runs up to 5 faster while retaining 0.989 full-bag utility.

Abstract

Whole-Slide Image Analysis

Explore similar work

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

Spatial Blindness in Whole-Slide Multiple Instance Learning

Mixture of Mini Experts: Overcoming the Linear Layer Bottleneck in Multiple Instance Learning