Multiple Instance Learning (MIL) has emerged as a promising paradigm for Whole Slide Image (WSI) diagnosis, offering effective learning with limited annotations. However, existing MIL frameworks overlook diagnostic priorities and fail to differentiate the severity of misclassifications in multiclass, leaving clinically critical errors unaddressed. We propose a mistake-severity-aware training strategy that organizes diagnostic classes into a hierarchical structure, with each level optimized using a severity-weighted cross-entropy loss that penalizes high-severity misclassifications more strongly. Additionally, hierarchical consistency is enforced through probabilistic alignment, a semantic feature remix applied to the instance bag to robustly train class priority and accommodate clinical cases involving multiple symptoms. An asymmetric Mikel's Wheel-based metric is also introduced to quantify the severity of errors specific to medical fields. Experiments on challenging public and real-world in-house datasets demonstrate that our approach significantly mitigates critical errors in MIL diagnosis compared to existing methods. We present additional experimental results on natural domain data to demonstrate the generalizability of our proposed method beyond medical contexts.
Figures & tables
Figure 1: The label characteristics that multiclass WSI have and the risks of applying MIL in the medical domain without accounting for mistake severity. (a) While every object in the natural domain is typically labeled, WSI only assigns the most urgent diagnosis among all observed complex findings. (b) Let ≺ denotes more urgent diagnosis. Despite the model MIL-A being considered a traditionally ”better” model than MIL-B because of high accuracy, it exhibits a higher number of severe misclassifications, which are shown in colored squares . This comparison highlights that conventional MIL approaches can lead to unreliable assessments for medical applications.
Figure 2: Structured class relationships from the finest hierarchy H to the root R .
Figure 3: Conceptual Descriptions of SFR. Dashed figures represent unique symptoms in Ya , where all figures are unlabeled and unknown. SFR extracts target cases and generates synthetic samples using only the available label Ya .
Figure 4: Two types of Mikel’s Wheel, where ∙ is the true class and ▶ is the predicted class. (a) In a symmetric wheel, the distance between two classes i and j is the absolute difference. (b) Proposed asymmetric wheel imposes a penalty P for misclassifying a high-priority true class.
BRACS [ 5 ]
In-house
TransMIL [ 35 ]
DTFD-MIL-AFS [ 42 ]
TransMIL [ 35 ]
DTFD-MIL-AFS [ 42 ]
ACC
AUC
AsCC
AsMC
ACC
AUC
AsCC
AsMC
ACC
AUC
AsCC
AsMC
ACC
AUC
AsCC
AsMC
Cross Entropy (CE)
40.23 ± 5.54
74.90 ± 3.26
58.48 ± 4.12
50.18 ± 2.99
45.98 ± 0.94
82.03 ± 1.92
61.98 ± 1.43
48.07 ± 3.54
90.41 ± 0.81
99.16 ± 0.01
93.13 ± 0.79
44.16 ± 5.92
91.45 ± 0.49
99.33 ± 0.06
93.71 ± 0.30
39.10 ± 1.65
Weighted CE (2:3:5)
42.15 ± 4.82
78.15 ± 3.58
59.22 ± 3.68
47.90 ± 0.76
47.51 ± 1.95
81.57 ± 1.63
63.16 ± 2.50
47.91 ± 4.78
91.88 ± 0.41
99.41 ± 0.04
94.27 ± 0.41
46.32 ± 3.90
91.65 ± 1.23
99.33 ± 0.09
93.99 ± 0.91
43.55 ± 2.52
Weighted CE (1:2:7)
41.76 ± 1.43
76.37 ± 0.90
58.96 ± 1.96
46.92 ± 3.85
45.21 ± 1.08
80.23 ± 0.16
62.77 ± 1.08
52.69 ± 3.34
89.17 ± 0.85
99.14 ± 0.08
92.29 ± 0.61
44.60 ± 0.52
91.57 ± 0.45
99.34 ± 0.05
93.95 ± 0.24
42.99 ± 3.79
Chang et al. [ 6 ]
47.51 ± 2.17
79.48 ± 1.55
63.98 ± 1.81
51.02 ± 1.66
45.21 ± 4.82
81.80 ± 2.21
61.31 ± 4.14
47.77 ± 2.55
91.42 ± 0.41
99.27 ± 0.02
94.01 ± 0.06
49.07 ± 1.14
91.07 ± 0.98
99.29 ± 0.10
93.61 ± 0.83
44.67 ± 0.41
Table 1: Classification results on various mistake severity comparison methods and our proposed method. The best performance is denoted in bold, and the second best is underlined.
Figure 5: Data class hierarchy diagram of In-house and BRACS [ 5 ] dataset. The In-house is presented in an italic font, whereas BRACS is underlined. We denote the class priority in each hierarchy using ≺ and ≡ .
Figure 6: Performance comparison on 182 mixed-symptom cases. Recall ‡ is measured by designating Adenoma (i.e., TA, TVA, TSA) as the positive class.
TransMIL [ 35 ]
Ablation
ACC
AUC
AsCC
AsMC
w.o. LMSCE
45.59 -2.46
78.79 -1.54
61.88 -2.52
47.10 -4.84
w.o. LHA
45.21 -2.84
79.61 -0.72
62.14 -2.26
51.41 -0.53
w.o. SFR
47.51 -0.54
76.88 -3.45
62.99 -1.41
47.92 -4.02
Ablated All
40.23 -7.82
74.90 -5.43
58.48 -5.92
50.18 -1.76
DTFD-MIL-AFS [ 42 ]
Table 2: Ablation results for both models on the BRACS dataset. The red values denote the difference in metrics between the fully-equipped model and the ablated results.
Figure 7: Visualization of various instance remix strategies. The green polygon indicates the most severe diagnosis labeled by the experts. Red boxes highlight the patches selected by the remix approach to synthesize a new sample without pixel-level annotation.
BRACS [ 5 ]
In-house
ACC
AsMC
ACC
AsMC
-
40.23 ± 5.54
50.18 ± 2.29
90.41 ± 0.81
44.16 ± 5.92
ReMix [ 41 ]
40.61 ± 4.63
52.37 ± 4.45
90.84 ± 0.62
51.64 ± 1.66
PseudoMix [ 26 ]
43.30 ± 1.43
51.07 ± 0.40
89.25 ± 0.45
51.64 ± 2.99
RandomMix [ 17 ]
44.06 ± 1.08
53.65 ± 4.45
89.25 ± 1.75
54.34 ± 2.22
SFR (Ours)
45.59 ± 3.55
53.18 ± 1.78
91.34 ± 0.09
55.86 ± 0.52
Table 3: Various instance remix strategies and their performance.
Figure 8: The confusion matrix simulates the predictions of the two models, with severe errors highlighted by color. The more in-depth analysis and computation of E(Risk) are detailed in the supplementary material.
ACC
AUC
AsCC
AsMC
Cross Entropy (CE)
83.24 ± 0.20
98.57 ± 0.04
87.23 ± 0.18
34.84 ± 0.77
Weighted CE (1:10)
73.69 ± 0.15
95.90 ± 0.16
79.79 ± 0.13
33.55 ± 0.21
CO2 [ 2 ]
83.97 ± 0.18
98.55 ± 0.03
87.77 ± 0.15
34.73 ± 0.29
CDW-CE [ 32 ]
84.11 ± 0.24
98.57 ± 0.01
87.87 ± 0.16
34.63 ± 0.40
MSCE (Ours)
85.64 ± 0.22
98.80 ± 0.02
89.12 ± 0.17
35.70 ± 0.61
Table 4: CIFAR-10 performance on varietal entropy terms.
Figure 9: Visualization of the expected class for various misclassified predictions. The values on the horizontal axis are the class index.
In-house
TA
TVA
TSA
HP
SSL
IP
LP
∑
Train
950
343
387
602
282
181
445
3,190
Validation
205
74
82
127
59
37
94
678
Test
300(95)
81(6)
102(18)
136(8)
114(55)
38
95
866(182)
BRACS [ 5 ]
IC
DCIS
FEA
ADH
UDH
PB
N
∑
Table 5: Data distribution over the classes. The values in parentheses represent 182 mixed cases, which were curated by the pathology experts.
Figure 10: Performance plot according to λ1,λ2 , and α .
Figure 11: Performance plot according to L and k .
Figure 12: Performance according to the refinement iterations T .
Figure 13: Ablation results for both models on the In-house dataset. The red values denote the difference in metrics between the fully-equipped model and the ablated results.
Method
Time Complexity
Mean Implementation Time (s) / Sample
BRACS [ 5 ]
In-house
ReMix [ 41 ]
O(L⋅k⋅T⋅D⋅n(Xi))
0.233
0.061
PseudoMix [ 26 ]
O(L⋅k⋅T⋅n(Xi))
0.015
0.004
RandomMix [ 17 ]
O(1)
0.005
0.003
SFR (Ours)
O(L⋅k⋅T⋅n(Xi)logn(Xi))
0.023
0.005
Table 6: Time complexity of various remix methods and their practical implementation time on the training set of the two datasets.
Figure 14: Additional visualization results on various remix strategies. The green polygon indicates the most severe diagnosis labeled by the experts. Red boxes highlight the patches selected by the remix methodology to synthesize the sample without pixel-level annotation.
Figure 15: Visualization of failure cases. The green polygon indicates the most severe diagnosis labeled by the experts. Red boxes highlight the patches selected by the remix methodology to synthesize the sample without pixel-level annotation.
Method
Equation
Distance Penalty
Asymmetric Penalty
Cross Entropy (CE)
−∑cY~[c]logp^[c]
✗
✗
Weighted CE
−∑cY~[c]wclogp^[c]
✗
✗
HXE [ 4 ]
−∑hexp(−αh(Nh))logp(Nh∣Nh+1) , where h(N) is depth of node N , p(Nh∣Nh+1)=∑B∈Leaves(Nh+1)p(B)∑A∈Leaves(Nh)p(A) , and p(N)=∏hp(Nh∣Nh+1)
△
✗
CO2 [ 2 ]
−∑cY~[c]logp^[c]+λ∑c1(c≥c′)×ReLU(δ+p^[c+1]−p^[c])+λ∑c1(c≤c′)×ReLU(δ+p^[c]−p^[c+1]) , where c′=argmaxcY~
✓
✗
CDW-CE [ 32 ]
−∑c(log(1−p^[c])×∣c−c′∣α) , , where c′=argmaxcY~
✓
✗
MSCE (Ours)
−p^WY⊤∑cY~[c]logp^[c]
✓
✓
Table 7: The Cross Entropy and its associated severity loss terms. Y~∈R1×C and p^∈R1×C indicate one-hot vectorized true label and predicted probability, respectively.
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.
Xiangyu Li, Ran Su
College of Intelligence and Computing Tianjin University
Whole-slide images may contain coexisting lesions, yet weak supervision typically provides only the most severe diagnosis. We formulate this setting as dominant-label multiple instance learning, where the observed label is the maximum-severity element of a latent lesion set. Standard multiclass MIL treats it as a complete one-hot description and therefore suppresses lower-severity evidence whenever it co-occurs with a dominant lesion. We propose DomSet-MIL, a label-aware framework that aligns the annotation mechanism across bag and instance levels. Dominant-Label Censored Likelihood models the reported diagnosis without declaring unreported lesions absent. Severity-Ordered Residual Routing assigns patch evidence from high to low severity and preserves residual support for secondary findings, while top-q pooling links local routing scores to bag prediction. Experiments across natural and controlled settings show that DomSet-MIL achieves state-of-the-art performance in dominant-label prediction while more effectively recovering latent coexisting-lesion evidence. These findings demonstrate that modeling the dominant-label mechanism can improve diagnostic prediction without sacrificing clinically meaningful secondary findings.
Lubin Gan, Jing Zhang, Xinming Zhang +3
USTC, Anhui, China · Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing, Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Anhui, China
Multiple instance learning (MIL) is widely used in computational pathology because it enables weakly supervised analysis of whole-slide images (WSIs) without requiring patch-level annotations. In attention-based MIL, instances with high attention scores are often interpreted as diagnostically important regions and used as visual explanations. However, attention scores alone cannot determine whether selected high-attention instances are significantly different from normal instances, limiting the reliability of attention-based explanations. In this paper, we formulate the evaluation of high-attention instances as a statistical hypothesis testing problem. Specifically, we assess whether a selected high-attention instance significantly deviates from a representative normal reference instance selected based on feature similarity. A major challenge is that both the target instance and the reference instance are selected through data-dependent procedures, rendering standard hypothesis testing invalid. To address this issue, we introduce a selective inference (SI) framework that explicitly accounts for the selection events induced by attention-based instance selection and adaptive reference selection, thereby enabling the computation of valid selective p-values conditional on these events. Experiments demonstrate Type-I error control on synthetic and MNIST-based data and practical applicability to pathological WSIs, with higher statistical power than the conventional over-conditioning approach.