Organizations: 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. · HKUST Shenzhen-Hong Kong Collaborative Innovation Research Institute, Futian, Shenzhen, China. · State Key Laboratory of Nervous System Disorders, The Hong Kong University of Science and Technology, Hong Kong, China. · SIAT-HKUST Joint Laboratory for Brain Science, The Hong Kong University of Science and Technology, Hong Kong, China.
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× 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.
Figures & tables
Figure 1: Overview of DI-MIL. (a) A whole-slide cytopathology image is partitioned into source tiles. (b) A frozen transformer-based foundation model extracts a global [CLS] representation and dense patch-token features from each tile. (c) DI-MIL performs intra-tile spherical clustering on dense token representations, aggregates tokens within each cluster, and constructs multiple decomposed instances per tile. (d) The resulting decomposed instances from all tiles are assembled into the slide-level MIL bag and processed by a downstream MIL aggregator.
Dataset
Instance
ABMIL
ACMIL
AUC
ACC
F1
AUC
ACC
F1
DINOv3
Cervical
Tile
91.40 (2.11)
84.56 (3.45)
84.40 (3.44)
93.83 (1.96)
87.22 (3.33)
87.01 (3.43)
DI (Ours)
91.52 (2.65)
85.69 (3.46)
85.48 (3.59)
93.86 (1.97)
88.24 (3.12)
88.06 (3.23)
Δ
+0.12
+1.13
+1.08
+0.03
+1.02
+1.05
Serous Fluid
Tile
85.37 (2.57)
78.76 (2.23)
78.70 (2.27)
86.11 (3.49)
79.13 (3.21)
79.11 (3.21)
Table 1: Comparison between Tile instances and our Decomposed instances (DI) across 4 cytopathology datasets, 3 foundation-model backbones, and 2 MIL aggregators. All methods use identical downstream MIL architectures and training protocols. Results are reported as mean with standard deviation in parentheses. Δ denotes the absolute performance change from Tile to DI.
Method
Cervical
Serous Fluid
Respiratory
Oral2026
AUC
ACC
F1
AUC
ACC
F1
AUC
ACC
F1
AUC
ACC
F1
DINOv3
ABMIL
91.40
84.56
84.40
85.37
78.76
78.70
79.38
74.73
70.64
88.06
79.84
70.93
ACMIL
93.83
87.22
87.01
86.11
79.13
79.11
81.05
75.98
72.11
88.37
80.77
71.45
DTFD
92.29
86.00
85.66
85.13
78.06
77.99
75.77
71.57
66.90
78.98
81.50
58.29
ASMIL
92.57
87.02
86.89
84.31
77.40
77.32
73.11
74.54
69.95
85.02
80.65
58.73
Table 2: Comparison with existing MIL methods across four cytopathology datasets and three foundation-model backbones. All methods within each backbone use identical features, data splits, and training protocols. DI-ACMIL denotes ACMIL operating on our proposed decomposed instances. Best results are shown in bold , and second-best results are underlined .
k
Oral2026
Respiratory
AUC
ACC
F1
AUC
ACC
F1
1
88.68 (3.93)
83.98 (3.03)
75.19 (4.25)
84.08 (2.16)
78.10 (2.97)
74.51 (3.34)
2
88.86 (4.22)
84.28 (4.43)
75.17 (6.72)
84.94 (2.85)
77.43 (3.34)
74.01 (4.01)
3
88.79 (5.35)
84.00 (6.63)
75.57 (9.55)
87.00 (2.71)
79.25 (3.50)
76.08 (4.72)
4
89.13 (4.86)
84.52 (5.25)
75.69 (8.27)
88.85 (2.07)
82.32 (3.10)
80.08 (3.48)
5
87.68 (6.23)
85.06 (5.35)
74.81 (11.59)
89.32 (2.35)
82.03 (4.64)
79.18 (6.05)
Table 3: Configuration sweep over the number of decomposed instances k on Oral2026 and Respiratory using Smart-CCS and ACMIL. All configurations use identical data splits and downstream training settings within each dataset and differ only in k . Results are reported as mean with standard deviation in parentheses.
Strategy
Oral2026
Respiratory
AUC
ACC
F1
AUC
ACC
F1
Random token grouping ( k=4 )
88.07 (4.41)
83.75 (3.50)
74.79 (5.68)
83.28 (2.42)
76.95 (2.78)
73.60 (3.85)
Spatial 2×2 grouping
86.83 (5.56)
83.38 (4.88)
73.95 (7.29)
86.11 (2.22)
80.11 (1.93)
77.15 (2.61)
DI ( k=4 )
89.13 (4.86)
84.52 (5.25)
75.69 (8.27)
88.85 (2.07)
82.32 (3.10)
80.08 (3.48)
Table 4: Comparison of alternative dense-token grouping strategies on Oral2026 and Respiratory using Smart-CCS and ACMIL. All strategies construct k=4 instances from the same dense-token representations and use identical downstream training settings. Results are reported as mean with standard deviation in parentheses.
Strategy
Oral2026
Respiratory
AUC
ACC
F1
AUC
ACC
F1
Tile 2048 w/o. DI
88.20 (5.44)
81.73 (6.14)
72.26 (4.25)
–
–
–
Tile 1200 w/o. DI
88.95 (5.06)
84.26 (4.78)
73.89 (4.73)
81.21 (3.45)
76.17 (3.19)
71.83 (4.47)
Tile 600 w/o. DI
88.73 (3.83)
84.24 (5.83)
74.40 (4.75)
83.81 (3.92)
78.10 (3.39)
74.77 (3.22)
Tile 300 w/o. DI
86.20 (3.34)
84.10 (4.88)
74.35 (4.84)
85.49 (2.11)
78.29 (2.93)
75.07 (3.78)
DI ( k=4 )
89.13 (4.86)
84.52 (5.25)
75.69 (8.27)
88.85 (2.07)
82.32 (3.10)
80.08 (3.48)
Table 5: Comparison of different tile-size strategies with our decomposed-instance (DI) strategy on Oral2026 and Respiratory . Tile n stands for using pixel tile size n without DI. All methods use Smart-CCS features, ACMIL, and the same downstream training protocol; DI uses the main-experiment setting k=4 with the dataset-specific source tile size (2048 on Oral2026 and 1200 on Respiratory ). Results are reported as mean with standard deviation in parentheses.
Figure 2: Performance and computational cost comparison between source tile size reduction and DI-MIL on Oral2026 using Smart-CCS with ACMIL. Left: AUC and macro-F1 versus the relative number of image tiles. Right: Relative tile count versus MIL feature-instance count.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Specimen
Slides
Task
Positive rate
Tile size
CV
Avg. tiles / slide
Cervical
Cervical
978
Negative vs. Positive
53.3%
1200 px
10-fold
1940
Serous Fluid
Serous fluid
2,717
Negative vs. Positive
51.5%
1200 px
10-fold
1572
Respiratory
Respiratory
1,041
Negative vs. Positive
66.3%
1200 px
10-fold
1141
Oral2026
Oral
363
Non-cancerous vs. Cancerous
17.6%
2048 px
5-fold
2579
Appendix
Table 6: Dataset and evaluation settings used in the experiments. Positive rate denotes the positive/cancerous class fraction. Avg. tiles / slide denotes the mean number of retained source tiles per WSI after background filtering.
Backbone
Encoder patch size (px)
Token grid
Dense tokens / encoded tile
Smart-CCS
14
16 × 16
256
DINOv3 (vit-l16)
16
14 × 14
196
UniCAS
16
14 × 14
196
Appendix
Table 7: Details of the backbones used in the experiments.
Setting
Value
Learning rate
2×10−4
Training seed
1024
Optimizer
Adam (ABMIL); AdamW (ACMIL)
Objective
CE (ABMIL); sub-token CE + slide CE + attention-diversity loss (ACMIL)
Default decomposition
k=4 ; 8 clustering iterations
Early stopping / maximum epochs
Enabled / 100
Appendix
Table 8: Core training and decomposition settings. Backend-specific optimizer/objective choices follow the corresponding official MIL formulation.
Figure 3: Mean random-corrected insertion-recall gain across 18 public cell-level datasets as a function of retained token budget. Higher values indicate that annotated cellular evidence is recovered more readily from the frozen dense-token representation.
Dataset
Smart-CCS
UniCAS
DINOv3
S−D
U−D
Test objects
APACS23 ( Harangi et al., 2024 )
0.1023
0.1045
0.0983
+0.0040
+0.0062
8,671
APC ( Cuña Cabrera, 2024 )
0.2263
0.2112
0.1569
+0.0694
+0.0542
281
Ascites2020 ( Su et al., 2020 )
0.4293
0.4213
0.3499
+0.0794
+0.0714
581
BCCD ( Mooney, 2018 )
0.1279
0.1279
0.1286
-0.0007
-0.0007
71
BTTFA ( Zhang, 2019 )
0.4551
0.4015
0.3914
+0.0637
+0.0101
978
CBC ( Alam and Islam, 2019 )
0.0532
0.0485
0.0548
-0.0016
-0.0063
61
Appendix
Table 9: Dataset-level random-corrected insertion AUC gain G for the three frozen backbones. Higher values indicate that annotated cellular evidence is recovered more readily from the dense-token representation. S−D and U−D denote the paired differences of Smart-CCS and UniCAS relative to DINOv3, respectively.
Figure 4: Visualization of dense token representations and decomposed instances. Patch tokens within the same color group form one DI. The color difference is positively correlated to feature vector distance in the encoder feature space.
Oral2026
Method
AUC
ACC
F1
[CLS]
88.20 (5.44)
81.73 (6.14)
72.26 (4.25)
DI ( k=1 )
88.68 (3.93)
83.98 (3.03)
75.19 (4.25)
DI ( k=4 )
89.13 (4.86)
84.52 (5.25)
75.69 (8.27)
Appendix
Table 10: Comparison of the conventional tile-level [CLS] representation, dense-token mean pooling ( k=1 ), and the default decomposed-instance setting ( k=4 ) across all four datasets. All variants use Smart-CCS, ACMIL, identical dataset-native source tiles, and the same downstream training protocol.
Figure 5: Qualitative comparison of high-attention instances across the four datasets. Row (i) shows conventional tile instances, whereas row (ii) shows decomposed instances, with red contours marking regions assigned to high-attention DIs. These examples are illustrative and do not constitute quantitative localization evaluation.
ABMIL
Dataset
n
Δ ( SDΔ )
95% CI
pW / pt
n+
DINOv3
Cervical
10
+0.12 (1.66)
[−1.07,1.31]
0.695 / 0.826
6/10
Serous Fluid
10
+0.42 (2.12)
[−1.09,1.93]
0.695 / 0.545
6/10
Respiratory
10
+1.38 (4.61)
[−1.92,4.68]
0.695 / 0.368
5/10
Oral2026
5
-2.34 (1.92)
[−4.72,0.04]
0.062 / 0.053
0/5
Appendix
Table 11: Fold-paired tests of Decomposed instances (DI) versus Tile on AUC. Each comparison uses the same folds, features, splits, and MIL protocol as Table 1 . Δ denotes the mean fold-wise difference DI − Tile in percentage points, and values in parentheses are SDΔ , the standard deviation of the paired differences. The 95% CI is the paired Student’s t interval for the mean Δ . pW is the two-sided exact Wilcoxon signed-rank p -value (zero differences dropped), and pt is the two-sided paired t -test p -value. n+ counts folds with Δ>0 . The three institutional cohorts use 10 folds, whereas Oral2026 uses 5 folds. For Oral2026 , the smallest attainable two-sided exact Wilcoxon p -value is 0.0625 when all five paired differences are non-zero.
Dataset
MIL
0
1
2
3
4
5
6
7
8
9
Med. [Q1,Q3]
DINOv3
Cervical
ABMIL
+0.55
+0.42
+0.44
-1.05
+2.26
-0.36
+2.34
-2.81
-1.79
+1.20
+0.43[−0.87,+1.03]
ACMIL
-3.85
-1.90
+4.31
+1.59
-3.22
0.00
+3.14
+2.05
-1.88
+0.06
+0.03[−1.90,+1.94]
Serous Fluid
ABMIL
+0.49
+3.64
-2.13
+1.03
+0.15
+4.19
-1.73
-1.49
+0.33
-0.27
+0.24[−1.19,+0.89]
ACMIL
+3.92
+4.11
-2.16
+3.51
-1.27
+2.07
-4.23
-1.87
+1.11
+1.65
+1.38[−1.72,+3.15]
Respiratory
ABMIL
-2.58
-2.69
+2.53
-0.29
+9.15
+0.25
+5.92
+7.16
-4.31
-1.30
−0.02[−2.26,+5.07]
Appendix
Table 12: Per-fold AUC Δ (DI − Tile, percentage points) for every matched comparison in Table 1 . The fold is the pairing unit. Med. [Q1,Q3] summarises the fold-wise Δ distribution.
Dataset
Metric
ABMIL
ACMIL
Mean Δ
95% CI
pW
Mean Δ
95% CI
pW
DINOv3
Cervical
ACC
+1.13
[−1.64,3.89]
0.301
+1.02
[−0.22,2.26]
0.109
F1
+1.08
[−1.72,3.89]
0.322
+1.05
[−0.21,2.31]
0.064
Serous Fluid
ACC
+0.77
[−0.96,2.50]
0.275
+0.96
[−1.48,3.39]
0.426
F1
+0.79
[−0.97,2.56]
0.232
+0.95
[−1.49,3.39]
0.557
Appendix
Table 13: Fold-paired tests of DI versus Tile on ACC and macro-F1, using the same matched folds as Table 1 . ACC and macro-F1 are treated as secondary metrics and are reported for completeness; their fold-paired analyses should not be interpreted as independent replications of the AUC analysis. Mean Δ , 95% paired- t CI, and two-sided exact Wilcoxon pW follow Table 11 . ∗p<0.05 , ∗∗p<0.01 , ∗∗∗p<0.001 .
Method
Cervical
Serous Fluid
AUC
ACC
F1
AUC
ACC
F1
DINOv3
ABMIL
91.40 (2.11)
84.56 (3.45)
84.40 (3.44)
85.37 (2.57)
78.76 (2.23)
78.70 (2.27)
ACMIL
93.83 (1.96)
87.22 (3.33)
87.01 (3.43)
86.11 (3.49)
79.13 (3.21)
79.11 (3.21)
DTFD
92.29 (2.72)
86.00 (3.49)
85.66 (3.68)
85.13 (3.13)
78.06 (2.42)
77.99 (2.41)
ASMIL
92.57 (1.79)
87.02 (3.33)
86.89 (3.40)
84.31 (3.01)
77.40 (3.43)
77.32 (3.52)
Appendix
Table 14: Detailed comparison with existing MIL methods on the Cervical and Serous Fluid datasets. Results are reported as mean with standard deviation in parentheses. Best mean results are shown in bold , and second-best mean results are underlined .
Method
Respiratory
Oral2026
AUC
ACC
F1
AUC
ACC
F1
DINOv3
ABMIL
79.38 (4.18)
74.73 (4.14)
70.64 (5.31)
88.06 (5.07)
79.84 (2.85)
70.93 (5.04)
ACMIL
81.05 (4.66)
75.98 (3.15)
72.11 (3.87)
88.37 (5.15)
80.77 (2.89)
71.45 (5.90)
DTFD
75.77 (6.15)
71.57 (5.40)
66.90 (5.63)
78.98 (14.35)
81.50 (2.93)
58.29 (14.02)
ASMIL
73.11 (8.67)
74.54 (4.51)
69.95 (6.47)
85.02 (6.08)
80.65 (3.23)
58.73 (12.90)
Appendix
Table 15: Detailed comparison with existing MIL methods on the Respiratory and Oral2026 datasets. Results are reported as mean with standard deviation in parentheses. Best mean results are shown in bold , and second-best mean results are underlined .
Method
Instances
Storage (GiB)
AUC
Macro-F1
Tile 2048
0.94M
1.79
88.20
72.26
Tile 1200
2.67M
5.09
88.95
73.89
DI-MIL ( k=4 )
3.75M
7.14
89.13
75.69
Tile 600
10.40M
19.83
88.73
74.40
Tile 300
40.52M
77.30
86.20
74.35
Appendix
Table 16: Absolute MIL feature-storage requirements on Oral2026 using Smart-CCS and ACMIL. Feature storage accounts for the float16 embeddings actually consumed by the MIL aggregator, excluding file-format overhead and auxiliary metadata. DI-MIL uses the original 2048-pixel tiles with k=4 . Performance is reported as mean AUC and macro-F1 (%).
Figure 6: Feature-extraction cost, MIL bag length, and AUC on Respiratory. (a) AUC versus measured CCS feature-extraction GPU-hours for conventional tiles (dashed) and DI ( k=4 ). Marker annotations give the median MIL bag relative to Tile 1200. (b) The same methods, split into extraction cost and MIL bag length, both relative to Tile 1200. DI matches the 4× bag of Tile 600 while remaining at Tile-1200 extraction cost. n=1041 slides; MIL times are isolated ACMIL steps on a RTX 3090 GPU and are not taken from the contended training wall-clock.
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
Multiple instance learning (MIL) has become a standard paradigm for whole slide image (WSI) analysis in digital pathology, as it enables slide-level prediction without dense annotations. Existing MIL methods typically rely on exhaustive extraction and encoding of high-resolution patches. However, this practice suffers from two critical limitations in real-world clinical settings: it struggles to capture global visual cues at lower magnifications, and incurs substantial computational overhead due to the massive number of high-resolution patches per slide. To address these limitations, we propose an efficient low-resolution multiple instance learning (LRMIL) framework that transfers high-resolution knowledge to low-resolution representations. LRMIL adopts a two-stage distillation strategy. First, patch-level cross-resolution distillation aligns low-resolution patch embeddings with high-resolution representations. Second, slide-level knowledge distillation trains a low-resolution student MIL model under both slide-level supervision and teacher guidance. At inference time, LRMIL operates exclusively on low-resolution patches, substantially reducing data preprocessing and computational cost. Extensive experiments on multiple WSI benchmarks demonstrate that LRMIL consistently outperforms state-of-the-art MIL methods while achieving more efficient inference. These results highlight LRMIL as a practical and scalable solution for WSI analysis in clinical pathology.
Yonghan Shin, Won-Ki Jeong
Department of Computer Science and Engineering, Korea University, Seoul, Korea
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.
Linfeng Ye, Shayan Mohajer Hamidi, Zhixiang Chi +5
University of Toronto · Stanford University · Hokkaido University