ASMIL: Attention-Stabilized Multiple Instance Learning for Whole Slide Imaging

Linfeng Ye; Shayan Mohajer Hamidi; Zhixiang Chi; Guang Li; Mert Pilanci; Takahiro Ogawa; Miki Haseyama; Konstantinos N. Plataniotis

cs.CVMar 1, 2026

ASMIL: Attention-Stabilized Multiple Instance Learning for Whole Slide Imaging

Authors: Linfeng Ye, Shayan Mohajer Hamidi, Zhixiang Chi, Guang Li, Mert Pilanci, Takahiro Ogawa, Miki Haseyama, Konstantinos N. Plataniotis

Organizations: University of Toronto · Stanford University · Hokkaido University

Abstract

Attention-based multiple instance learning (MIL) has emerged as a powerful framework for whole slide image (WSI) diagnosis, leveraging attention to aggregate instance-level features into bag-level predictions. Despite this success, we find that such methods exhibit a new failure mode: unstable attention dynamics. Across four representative attention-based MIL methods and two public WSI datasets, we observe that attention distributions oscillate across epochs rather than converging to a consistent pattern, degrading performance. This instability adds to two previously reported challenges: overfitting and over-concentrated attention distribution. To simultaneously overcome these three limitations, we introduce attention-stabilized multiple instance learning (ASMIL), a novel unified framework. ASMIL uses an anchor model to stabilize attention, replaces softmax with a normalized sigmoid function in the anchor to prevent over-concentration, and applies token random dropping to mitigate overfitting. Extensive experiments demonstrate that ASMIL achieves up to a 6.49% F1 score improvement over state-of-the-art methods. Moreover, integrating the anchor model and normalized sigmoid into existing attention-based MIL methods consistently boosts their performance, with F1 score gains up to 10.73%. All code and data are publicly available at https://github.com/Linfeng-Ye/ASMIL.

Figures & tables

Figure 1: Visualization of attention dynamics on a tumor WSI for TransMIL ( Shao et al., 2021 ) vs. ASMIL (our method). The green contours in the figures indicate the annotated tumor regions. Top : TransMIL attention distribution at selected training iterations. Middle : Jensen-Shannon divergence (JSD) between attention distributions at successive steps and the cross entropy loss (CE), comparing TransMIL (blue) and ASMIL (red). Bottom : Attention distribution from ASMIL over different training iterations. Due to the weakly supervised nature of WSI subtyping datasets, TransMIL’s attention patterns never converge during training, further, it focuses on only a subset of cancerous regions. In contrast, our method ( $i$ ) produces stable attention distributions throughout training and ( $ii$ ) consistently highlights cancerous regions.

Figure 2: Overview of ASMIL. Each WSI is divided into tiles and embedded into vision tokens using a pretrained encoder. These tokens, along with trainable FEAT tokens, feed into both online and anchor encoders. The anchor encoder’s attention scores over the FEAT tokens are transformed into a probability vector using an NSF, while the online encoder applies a softmax. To stabilize training and prevent the online model’s attention from becoming overly concentrated, we compute the KL divergence between the two distributions. Gradients are blocked to the anchor encoder using a stop-gradient (sg) operator, and its parameters are updated via EMA from the online encoder. During training, we randomly drop (rd) $N$ FEAT tokens, feed the remaining tokens into a second transformer with a trainable [CLS] token, and train a classifier on its output. and indicate learnable and frozen components, respectively.

Figure 3: (a) Distribution of attention scores in ABMIL, which exhibits a long-tailed pattern. (b) Attention distribution obtained with the softmax function and (c) with the NSF. Unlike softmax, the normalized sigmoid suppresses large values in the long tail, yielding a less sparse and more interpretable attention distribution.

	Dataset	CAMELYON-16		CAMELYON-17		BRACS
Backbone	Method	F1 score $\uparrow$	AUC $\uparrow$	F1 score $\uparrow$	AUC $\uparrow$	F1 score $\uparrow$	AUC $\uparrow$
ResNet-18 ImageNet Pretrained	ABMIL ICML 2018	$0.757\scriptscriptstyle\pm 0.020$	$0.790\scriptscriptstyle\pm 0.027$	$0.508\scriptscriptstyle\pm 0.032$	$0.779\scriptscriptstyle\pm 0.021$	$0.523\scriptscriptstyle\pm 0.028$	$0.723\scriptscriptstyle\pm 0.035$
	Clam-SB Nature 2021	$0.742\scriptscriptstyle\pm 0.024$	$0.763\scriptscriptstyle\pm 0.049$	$0.504\scriptscriptstyle\pm 0.012$	$0.778\scriptscriptstyle\pm 0.024$	$0.521\scriptscriptstyle\pm 0.046$	$0.750\scriptscriptstyle\pm 0.039$
	TransMIL NeurIPS 2021	$0.643\scriptscriptstyle\pm 0.088$	$0.706\scriptscriptstyle\pm 0.076$	$0.499\scriptscriptstyle\pm 0.082$	$0.794\scriptscriptstyle\pm 0.053$	$0.444\scriptscriptstyle\pm 0.040$	$0.732\scriptscriptstyle\pm 0.043$
	DSMIL CVPR 2021b	$0.736\scriptscriptstyle\pm 0.025$	$0.773\scriptscriptstyle\pm 0.034$	$0.473\scriptscriptstyle\pm 0.052$	$0.705\scriptscriptstyle\pm 0.022$	$0.511\scriptscriptstyle\pm 0.052$	$0.751\scriptscriptstyle\pm 0.028$
	DTFD-MIL CVPR 2022	$0.758\scriptscriptstyle\pm 0.051$	$0.815\scriptscriptstyle\pm 0.063$	$0.546\scriptscriptstyle\pm 0.010$	$0.735\scriptscriptstyle\pm 0.011$	$0.469\scriptscriptstyle\pm 0.016$	$0.717\scriptscriptstyle\pm 0.032$

Table 1: The F1 score and AUC of different MIL approaches across three WSI datasets. Bold and underlined values denote the best and second-best results, respectively.

Dataset			CAMELYON-16		CAMELYON-17		BRACS
Method	Anchor	NSF	F1 score $\uparrow$	AUC $\uparrow$	F1 score $\uparrow$	AUC $\uparrow$	F1 score $\uparrow$	AUC $\uparrow$
ABMIL ICML 2018	✗	✗	$0.914\scriptscriptstyle\pm 0.031$	$0.945\scriptscriptstyle\pm 0.027$	$0.522\scriptscriptstyle\pm 0.050$	$0.853\scriptscriptstyle\pm 0.016$	$0.680\scriptscriptstyle\pm 0.051$	$0.866\scriptscriptstyle\pm 0.029$
	✓	✗	$0.951\scriptscriptstyle\pm 0.015$	$0.963\scriptscriptstyle\pm 0.008$	$0.573\scriptscriptstyle\pm 0.011$	$0.871\scriptscriptstyle\pm 0.010$	$0.751\scriptscriptstyle\pm 0.013$	$0.877\scriptscriptstyle\pm 0.007$
			$+0.037$	$+0.018$	$+0.051$	$+0.018$	$+0.071$	$+0.011$
	✓	✓	$0.953\scriptscriptstyle\pm 0.009$	$0.967\scriptscriptstyle\pm 0.006$	$0.574\scriptscriptstyle\pm 0.010$	$0.883\scriptscriptstyle\pm 0.014$	$0.753\scriptscriptstyle\pm 0.009$	$0.887\scriptscriptstyle\pm 0.014$
			$+0.039$	$+0.022$	$+0.052$	$+0.030$	$+0.073$	$+0.021$

Table 2: Applying anchor model and NSF to other attention-based MIL methods.

Anchor	NSF	rd	F1 score $\uparrow$	AUC $\uparrow$
✓	✓	✓	$\mathbf{0.781\scriptscriptstyle\pm 0.042}$	$\mathbf{0.914\scriptscriptstyle\pm 0.014}$
✓	✓	✗	$0.765\scriptscriptstyle\pm 0.030$	$0.903\scriptscriptstyle\pm 0.018$
✓	✗	✓	$0.759\scriptscriptstyle\pm 0.028$	$0.895\scriptscriptstyle\pm 0.012$
✓	✗	✗	$0.747\scriptscriptstyle\pm 0.026$	$0.887\scriptscriptstyle\pm 0.015$
✗	✗	✓	$0.728\scriptscriptstyle\pm 0.019$	$0.868\scriptscriptstyle\pm 0.010$
✗	✗	✗	$0.712\scriptscriptstyle\pm 0.020$	$0.860\scriptscriptstyle\pm 0.012$

Table 3: Component-wise ablation of ASMIL on BRACS. We evaluate the contribution of the anchor model, NSF, and random drop (rd).

Figure 4: Visual comparison of attention maps on the CAMELYON-16 dataset. The left column shows the original WSI with ground-truth tumor annotations outlined in red; the remaining columns present attention maps for ASMIL (ours), DTFD-MIL, CAMIL, and TransMIL (left to right).

Appendix figures & tables38 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: T-SNE embeddings of ASMIL bag-level features on the BRACS training set across training epochs. Top: with the anchor model; Bottom: without the anchor model.

Figure 6: T-SNE embeddings of TransMIL bag-level features on the BRACS training set across training epochs. Top: with the anchor model; Bottom: without the anchor model.

Figure 7: Ablation study of applying the softmax function with temperature scaling to the attention scores: (a) attention distribution of the proposed ASMIL, (b) softmax with $T=2$ applied to the anchor model, (c) softmax with $T=4$ applied to the anchor model, (d) softmax with $T=8$ applied to the anchor model.

Figure 8: Histograms of (a) raw attention scores, (b) attention distribution obtained by the softmax function with temperature $T=1$ , (c) $T=2$ , (d) $T=4$ , (e) $T=8$ , and (f) attention distribution computed using an NSF . The Y-axis is displayed on a logarithmic scale for better visualization.

BRACS
Normalized Sigmoid		Softmax T=2		Softmax T=4		Softmax T=8
F1 score $\uparrow$	AUC $\uparrow$	F1 score $\uparrow$	AUC $\uparrow$	F1 score $\uparrow$	AUC $\uparrow$	F1 score $\uparrow$	AUC $\uparrow$
$0.781\scriptscriptstyle\pm 0.042$	$0.914\scriptscriptstyle\pm 0.014$	$0.667\scriptscriptstyle\pm 0.049$	$0.860\scriptscriptstyle\pm 0.027$	$0.712\scriptscriptstyle\pm 0.029$	$0.876\scriptscriptstyle\pm 0.012$	$0.688\scriptscriptstyle\pm 0.037$	$0.858\scriptscriptstyle\pm 0.031$

Appendix

Table 4: Subtyping performance on BRACS, when we apply softmax with temperature scaling to the anchor model.

CAMELYON-16
Metric	NSF	$\text{entmax}_{\alpha=2}$ (sparsemax)	$\text{entmax}_{\alpha=1.75}$	$\text{entmax}_{\alpha=1.5}$ (entmax-15)	$\text{entmax}_{\alpha=1.25}$	$\text{entmax}_{\alpha=1}$ (softmax)
F1 score $\uparrow$	$0.965\scriptscriptstyle\pm 0.020$	$0.938\scriptscriptstyle\pm 0.031$	$0.927\scriptscriptstyle\pm 0.034$	$0.937\scriptscriptstyle\pm 0.014$	$0.910\scriptscriptstyle\pm 0.026$	$0.942\scriptscriptstyle\pm 0.0147$
AUC $\uparrow$	$0.985\scriptscriptstyle\pm 0.017$	$0.964\scriptscriptstyle\pm 0.012$	$0.959\scriptscriptstyle\pm 0.017$	$0.960\scriptscriptstyle\pm 0.017$	$0.925\scriptscriptstyle\pm 0.031$	$0.963\scriptscriptstyle\pm 0.020$
Time per epoch $\downarrow$	6.340s	8.451s	8.452s	8.451s	8.450s	6.336s

Appendix

Table 5: ASMIL performance when replacing NSF with entmax on CAMELYON-16.

Dataset	CAMELYON-16
Online NSF	F1 score $\uparrow$	AUC $\uparrow$
✓	$0.920\scriptscriptstyle\pm 0.020$	$0.936\scriptscriptstyle\pm 0.021$
✗	$0.965\scriptscriptstyle\pm 0.020$	$0.985\scriptscriptstyle\pm 0.017$
Dataset	BRACS
Online NSF	F1 score $\uparrow$	AUC $\uparrow$
✓	$0.726\scriptscriptstyle\pm 0.014$	$0.865\scriptscriptstyle\pm 0.017$

Appendix

Table 6: Ablation study on the impact of applying the normalized sigmoid (NS) function to both the online and anchor models. ✓ indicates that NSF is applied to both models, while ✗ denotes the default setting where NSF is applied only to the anchor model. Subtyping performance is evaluated on the CAMELYON-16 and BRACS datasets using F1 score and AUC. A significant performance drop is observed on CAMELYON-16 when NSF is applied to both models.

Dataset	CAMELYON-16		CAMELYON-17		BRACS
Mertic	F1 score	AUC	F1 score	AUC	F1 score	AUC
ASMIL W. SoftMax	$0.965\scriptscriptstyle\pm 0.020$	$0.985\scriptscriptstyle\pm 0.017$	$0.689\scriptscriptstyle\pm 0.005$	$0.898\scriptscriptstyle\pm 0.010$	$0.781\scriptscriptstyle\pm 0.042$	$0.914\scriptscriptstyle\pm 0.014$
ASMIL W. Mixture	$0.953\scriptscriptstyle\pm 0.023$	$0.972\scriptscriptstyle\pm 0.030$	$0.686\scriptscriptstyle\pm 0.012$	$0.889\scriptscriptstyle\pm 0.009$	$0.774\scriptscriptstyle\pm 0.054$	$0.910\scriptscriptstyle\pm 0.067$
$\zeta$ in Equation 34	0.9952		0.9894		0.9963

Appendix

Table 7: Comparison between ASMIL trained with the standard softmax function in the online model (ASMIL w. Softmax) and with the mixed attention function defined in Equation 34 (ASMIL w. Mixture). The more flexible trainable mapping does not yield improvements over the simpler softmax baseline.

Dataset	CAMELYON-16	CAMELYON-17	BRACS
The ratio of affine dependent feature bags	99.24%	99.80%	96.08%

Appendix

Table 8: Ratio of affinely dependent feature bags in the CAMELYON-16, CAMELYON-17, and BRACS datasets; most bags are affinely dependent.

PathGen-Clip-VIT-L
Dataset	CAMELYON-16		CAMELYON-17
Method	F1 score $\uparrow$	AUC $\uparrow$	F1 score $\uparrow$	AUC $\uparrow$
Clam-SB	$0.941\scriptscriptstyle\pm 0.014$	$0.960\scriptscriptstyle\pm 0.015$	$0.622\scriptscriptstyle\pm 0.031$	$0.899\scriptscriptstyle\pm 0.012$
TransMIL	$0.951\scriptscriptstyle\pm 0.024$	$0.968\scriptscriptstyle\pm 0.028$	$0.656\scriptscriptstyle\pm 0.021$	$0.892\scriptscriptstyle\pm 0.014$
DSMIL	$0.895\scriptscriptstyle\pm 0.038$	$0.949\scriptscriptstyle\pm 0.017$	$0.582\scriptscriptstyle\pm 0.062$	$0.887\scriptscriptstyle\pm 0.013$
IBMIL	$0.935\scriptscriptstyle\pm 0.014$	$0.953\scriptscriptstyle\pm 0.009$	$0.629\scriptscriptstyle\pm 0.027$	$0.884\scriptscriptstyle\pm 0.016$

Appendix

Table 9: The F1 score and AUC of different MIL approaches on two WSI subtyping datasets.

Figure 9: Ablation study on the coefficient $\beta$ , on CAMELYON-16 dataset.

Figure 10: Ablation study on the coefficient

\beta

, on BRACS dataset.

CAMELYON-16
# FEAT tokens	2	4	8	16
F1 score $\uparrow$	$0.930\scriptscriptstyle\pm 0.012$	$0.946\scriptscriptstyle\pm 0.009$	$0.965\scriptscriptstyle\pm 0.012$	$0.960\scriptscriptstyle\pm 0.006$
AUC $\uparrow$	$0.932\scriptscriptstyle\pm 0.017$	$0.973\scriptscriptstyle\pm 0.011$	$0.985\scriptscriptstyle\pm 0.013$	$0.981\scriptscriptstyle\pm 0.009$
CAMELYON-17
F1 score $\uparrow$	$0.556\scriptscriptstyle\pm 0.012$	$0.610\scriptscriptstyle\pm 0.009$	$0.674\scriptscriptstyle\pm 0.016$	$0.689\scriptscriptstyle\pm 0.005$
AUC $\uparrow$	$0.784\scriptscriptstyle\pm 0.019$	$0.833\scriptscriptstyle\pm 0.011$	$0.879\scriptscriptstyle\pm 0.024$	$0.898\scriptscriptstyle\pm 0.010$

Appendix

Table 10: Ablation results on the number of tokens on different WSI datasets.

Dataset	BRACS
Update	F1 score $\uparrow$	AUC $\uparrow$
Epoch	$0.742\scriptscriptstyle\pm 0.015$	$0.871\scriptscriptstyle\pm 0.003$
Batch	$0.781\scriptscriptstyle\pm 0.042$	$0.914\scriptscriptstyle\pm 0.014$
Dataset	CAMELYON-16
Update	F1 score $\uparrow$	AUC $\uparrow$
Epoch	$0.920\scriptscriptstyle\pm 0.020$	$0.936\scriptscriptstyle\pm 0.021$

Appendix

Table 11: Ablation study on anchor model update frequency, where batch-wise updates consistently outperform epoch-wise updates in both F1 score and AUC on BRACS and CAMELYON-16.

Figure 11: Ablation study of random drop probability ( $B$ ) vs. model F1 score and AUC on CAMELYON-16 (top row) and BRACS (bottom row). Across both datasets and trainable-token settings (a) 4 tokens, (b) 8 tokens, and (c) 16 tokens, the test F1 score and AUC consistently peak around $B=0.5$ .

Panel of Figure 12 — Figure 12: Performance comparison between ABMIL ( Ilse et al., 2018 ) , ASMIL with random drop (ASMIL W. rd), and ASMIL without random drop (ASMIL w/o rd). Both ABMIL and ASMIL w/o rd show signs of overfitting, as their F1 score and AUC peak and then decline. In contrast, ASMIL with random drop maintains stable performance across training, demonstrating that random drop effectively mitigates overfitting.

Dataset	Model	AUC	F1	$p_{\text{AUC}}$	$p_{\text{F1}}$
CAMELYON-16	w/o anchor	0.942	0.979	0.013	0.024
CAMELYON-16	w/ anchor	0.967	0.983	0.013	0.024
CAMELYON-17	w/o anchor	0.642	0.879	0.024	0.035
CAMELYON-17	w/ anchor	0.693	0.899	0.024	0.035
BRACS	w/o anchor	0.729	0.866	0.012	0.009
BRACS	w/ anchor	0.784	0.916	0.012	0.009

Appendix

Table 12: Statistical comparison of ASMIL with and without the anchor model. We report the mean performance over

10

random seeds along with p-values from DeLong tests for AUC and permutation tests for F1.

Figure 13: Additional qualitative examples of tumor regions and ASMIL attention maps. Rows 1 and 3 show the ground-truth tumor annotations (cancerous regions outlined in red), and rows 2 and 4 show the corresponding ASMIL attention maps.

Method	Dice $\uparrow$	Specificity $\uparrow$	FROC $\uparrow$
CLAM-SB	$0.459$	$0.987$	$0.4257$
TransMIL	$0.103$	$0.999$	$0.0866$
DTFD-MIL	$0.525$	$0.999$	$0.4712$
DSMIL	$0.259$	$0.863$	$0.4506$
CAMIL	$0.515$	$0.980$	$0.4612$
ASMIL	$0.586$	$0.999$	$0.4941$

Appendix

Table 13: Localization results on CAMELYON-16.

BRACS	Training
Method	CLAM-SB	ABMIL	TransMIL	MHIM-MIL	ASMIL
Time	2.26s	0.95s	5.99s	19.4s	7.49s
Memory	94MB	90MB	340MB	2178MB	570MB
BRACS	Inference
FLOPs	162M	164M	781M	345M	542M
Time	0.45s	0.37s	0.74s	0.40s	0.52s

Appendix

Table 14: Computational cost on BRACS (lower is better). We report training time and peak memory per epoch, and inference FLOPs, latency, and memory. ASMIL (ours) delivers efficient inference, cutting compute by

30.6\%

, latency by

29.2\%

, and memory by

20.3\%

compared with TransMIL, while requiring

4\times

less training memory than MHIM-MIL.

BRACS	Training		Inference
Method	Time	Memory	FLOPs	Time	Memory
CLAM-SB w/o anchor	2.26s	94MB	162M	0.45s	69MB
CLAM-SB W. anchor	2.69s	120MB	162M	0.45s	69MB
TransMIL w/o anchor	5.99s	340MB	781M	0.74s	246MB
TransMIL W. anchor	7.27s	443MB	781M	0.74s	246MB
DSMIL w/o anchor	0.57s	60 MB	103M	1.09s	113 MB

Appendix

Table 15: Inference FLOPs, training time per epoch (Time), and memory usage (Memory) for four well-known methods, CLAM-SB, TransMIL, DSMIL, and ABMIL, with and without the anchor model. The anchor model incurs only minor computational overhead. FLOPs are measured using a fixed bag size of 2000 instances.

Method	BLCA	BRCA	GBMLGG	LUAD	UCEC
ABMIL ICML 2018	$0.5581\scriptscriptstyle\pm 0.031$	$0.5825\scriptscriptstyle\pm 0.035$	$0.7935\scriptscriptstyle\pm 0.032$	$0.6121\scriptscriptstyle\pm 0.050$	$0.6667\scriptscriptstyle\pm 0.033$
TransMIL NeurIPS 2021	$0.5885\scriptscriptstyle\pm 0.055$	$0.6140\scriptscriptstyle\pm 0.060$	$0.7956\scriptscriptstyle\pm 0.015$	$0.5708\scriptscriptstyle\pm 0.050$	$0.6380\scriptscriptstyle\pm 0.067$
ILRA ICLR 2023	$0.5549\scriptscriptstyle\pm 0.053$	$0.5705\scriptscriptstyle\pm 0.067$	$0.7742\scriptscriptstyle\pm 0.014$	$0.5179\scriptscriptstyle\pm 0.081$	$0.6503\scriptscriptstyle\pm 0.064$
$\text{R}^{2}\text{T-MIL }$ CVPR 2024	$0.5775\scriptscriptstyle\pm 0.024$	$0.5476\scriptscriptstyle\pm 0.095$	$0.7757\scriptscriptstyle\pm 0.024$	$0.5711\scriptscriptstyle\pm 0.076$	$0.6510\scriptscriptstyle\pm 0.087$
DeepAttnMISL MIA 2020	$0.5646\scriptscriptstyle\pm 0.035$	$0.5346\scriptscriptstyle\pm 0.036$	$0.6750\scriptscriptstyle\pm 0.048$	$0.4678\scriptscriptstyle\pm 0.039$	$0.6259\scriptscriptstyle\pm 0.086$
Patch-GCN MICCAI 2021	$0.6124\scriptscriptstyle\pm 0.031$	$0.6375\scriptscriptstyle\pm 0.033$	$0.7999\scriptscriptstyle\pm 0.021$	$0.5922\scriptscriptstyle\pm 0.053$	$0.7212\scriptscriptstyle\pm 0.025$

Appendix

Table 16: C-index for WSI-based survival prediction using vision-only MIL models.

Dataset	Domain	Bags (pos/neg)	Total instances	Dim./inst.
MUSK1	Drug activity	92 (47/45)	476	166
MUSK2	Drug activity	102 (39/63)	6598	166
TIGER	Images (Blobworld segments)	200 (100/100)	1220	230
FOX	Images (Blobworld segments)	200 (100/100)	1320	230
ELEPHANT	Images (Blobworld segments)	200 (100/100)	1391	230

Appendix

Table 17: MIL dataset statistics.

Methods	MUSK1	MUSK2	FOX	TIGER	ELEPHANT
ABMIL ICML 2018	$0.916\scriptscriptstyle\pm 0.118$	$0.928\scriptscriptstyle\pm 0.109$	$0.952\scriptscriptstyle\pm 0.051$	$0.953\scriptscriptstyle\pm 0.042$	$0.969\scriptscriptstyle\pm 0.036$
DSMIL CVPR 2021b	$0.959\scriptscriptstyle\pm 0.053$	$0.952\scriptscriptstyle\pm 0.066$	$0.939\scriptscriptstyle\pm 0.060$	$0.951\scriptscriptstyle\pm 0.053$	$\mathbf{0.989\scriptscriptstyle\pm 0.023}$
TransMIL NeurIPS 2021	$0.927\scriptscriptstyle\pm 0.093$	$0.877\scriptscriptstyle\pm 0.127$	$0.944\scriptscriptstyle\pm 0.050$	$0.963\scriptscriptstyle\pm 0.042$	$0.979\scriptscriptstyle\pm 0.030$
DEMIL NeurIPS 2023a	$0.963\scriptscriptstyle\pm 0.073$	$0.961\scriptscriptstyle\pm 0.057$	$0.941\scriptscriptstyle\pm 0.047$	$0.965\scriptscriptstyle\pm 0.035$	$0.969\scriptscriptstyle\pm 0.034$
RGMIL Neurips 2023	$0.968\scriptscriptstyle\pm 0.060$	$0.963\scriptscriptstyle\pm 0.048$	$0.954\scriptscriptstyle\pm 0.048$	$0.949\scriptscriptstyle\pm 0.047$	$0.965\scriptscriptstyle\pm 0.032$
PSMIL ICLR 2025	$0.968\scriptscriptstyle\pm 0.053$	$\mathbf{0.968\scriptscriptstyle\pm 0.052}$	$0.942\scriptscriptstyle\pm 0.054$	$0.947\scriptscriptstyle\pm 0.047$	$0.985\scriptscriptstyle\pm 0.030$

Appendix

Table 18: Results on the small MIL benchmark datasets.

Figure 14: Visualization of attention dynamics on a normal WSI for ABMIL vs. ABMIL + anchor + NSF.

Figure 15: Visualization of attention dynamics on a tumor WSI for ABMIL vs. ABMIL + anchor + NSF.

Figure 16: Visualization of attention dynamics on a normal WSI for TransMIL vs. TransMIL + anchor + NSF.

Figure 17: Visualization of attention dynamics on a tumor WSI for TransMIL vs. TransMIL + anchor + NSF.

Figure 18: Visualization of attention dynamics on a normal WSI for DSMIL vs. DSMIL + anchor + NSF.

Figure 19: Visualization of attention dynamics on a tumor WSI for DSMIL vs. DSMIL + anchor + NSF.

Figure 20: Visualization of attention dynamics on a normal WSI for ABMIL vs. ABMIL + anchor + NSF.

Figure 21: Visualization of attention dynamics on a tumor WSI for ABMIL vs. ABMIL + anchor + NSF.

Figure 22: Visualization of attention dynamics on a tumor WSI for TransMIL vs. TransMIL + anchor + NSF.

Figure 23: Visualization of attention dynamics on a tumor WSI for TransMIL vs. TransMIL + anchor + NSF.

Figure 24: Visualization of attention dynamics on a tumor WSI for DSMIL vs. DSMIL + anchor + NSF.

Figure 25: Visualization of attention dynamics on a tumor WSI for DSMIL vs. DSMIL + anchor + NSF.

Figure 26: Left: annotated WSI. Right: attention map generated by ASMIL, which fails to assign high attention to all tumor regions, as highlighted by the green arrow.

Method	Clam	TransMIL	DTFD-MIL	DSMIL	CAMIL	ASMIL
Rate	46.93%	3.246%	50.34%	48.22%	49.78%	54.63%

Appendix

Table 19: Rate of cancerous WSIs without missed regions on the CAMELYON-16 dataset.

Explore similar work

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
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
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

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

Explore similar work

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

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

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