cs.CVOct 21, 2026
SaveLEAF-Mamba: Feature Adaptation and Fusion for Whole Slide Image Analysis
Organizations: National Engineering Laboratory for Brain-inspired Intelligence Technology and Application, MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China, Hefei, 230026, China · Anhui Province Key Laboratory of Biomedical Imaging and Intelligent Processing, Institute of Artificial Intelligence, Hefei Comprehensive National Science Center, Hefei, 230088, China · School of Software Technology, Zhejiang University, Ningbo, 315048, China · Department of Biomedical Informatics, Harvard Medical School, Boston, MA, 02115, USA ·
Abstract
Whole slide image (WSI) analysis enables computational diagnosis and prognosis, but gigapixel images and limited annotations make slide-level learning challenging. Multiple instance learning addresses this task by aggregating patch representations from frozen pretrained encoders. Recently, Mamba has enabled efficient modeling of long patch sequences. However, task supervision alone does not explicitly constrain information recovery during feature adaptation, while Mamba couples local processing with long-range propagation rather than retaining separate representations for fusion. Motivated by these challenges, we introduce LEAF-Mamba, a framework that jointly addresses latent evidence adaptation and fusion for frozen-feature WSI analysis. A reconstruction-regularized adapter learns compact, task-specific representations while reconstructing the input features. Parallel Mamba and convolutional streams then capture long-range dependencies and sequence-local interactions, with an input-dependent gate modulating their joint representation before slide-level pooling. Experiments across feature extractors demonstrate improved average performance over strong MIL baselines in both diagnostic classification and survival prediction. Component analyses further support the contributions of adaptation and fusion. These findings highlight the value of jointly designing feature transformation and contextual aggregation to translate pretrained representations into effective slide-level predictions.
Figures & tables
Figure 1: Conceptual visualization of Mamba-guided evidence fusion.
Figure 2: Overview of LEAF-Mamba. Frozen features undergo reconstruction-regularized latent evidence adaptation and Mamba-guided fusion of long-range sequence context, short-range token interactions, and content-dependent gating for slide-level prediction.
Figure 3: Effect of LEA on BRACS features.
Figure 4: Diagnostic classification and TCGA survival prediction results across frozen feature extractors. The vertical axis shows the score difference from the best baseline using the same feature extractor, on a nonlinear scale.
| (a) BRACS and BRACS official BRACS BRACS official Method AUC ACC F1 AUC ACC F1 MeanMIL 69.9 41.4 29.0 65.2 37.0 22.8 MaxMIL 67.3 36.8 27.1 63.5 26.0 26.9 CLAM 79.3 44.4 37.5 73.2 37.0 32.3 ABMIL 75.9 43.2 34.7 71.5 33.0 26.1 MHIM-ABMIL 76.9 43.5 35.9 72.0 34.0 25.9 DSMIL 67.0 35.8 24.9 70.5 35.0 27.9 DTFD-MIL 76.9 41.3 34.3 70.9 30.0 24.3 TransMIL 73.0 39.5 25.8 66.4 32.0 18.9 MHIM-TransMIL 74.3 42.6 26.8 67.3 33.0 19.3 S4MIL 76.4 44.0 33.6 69.1 31.0 18.0 RRTMIL 78.2 44.1 34.0 68.4 31.0 19.6 MambaMIL 79.4 46.6 35.3 74.1 31.0 27.9 CAMIL 77.8 44.6 33.7 70.7 32.0 24.6 RPMIL 80.0 47.2 36.5 74.8 34.0 27.9 ACMIL 79.0 46.0 35.8 73.9 34.0 27.6 Jigsaw 80.4 47.6 36.9 75.0 35.0 28.2 LEAF-Mamba 81.6 49.2 37.3 76.9 35.0 28.9 Best scores in each column are in bold. |
| (b) Camelyon16 and EBRAINS Camelyon16 EBRAINS Method AUC ACC F1 AUC ACC F1 MeanMIL 70.9 65.2 55.4 91.0 37.3 34.6 MaxMIL 94.2 88.5 85.1 87.9 32.9 30.4 CLAM 93.5 89.5 86.6 91.8 37.3 34.8 ABMIL 94.1 87.7 84.4 91.7 39.7 37.6 MHIM-ABMIL 94.3 89.7 82.9 91.8 41.3 38.4 DSMIL 88.9 83.2 78.9 90.3 40.6 39.9 DTFD-MIL 93.8 88.7 85.9 89.5 35.2 33.5 TransMIL 92.2 81.7 79.4 92.7 38.2 35.8 MHIM-TransMIL 93.7 82.5 80.3 91.4 38.6 34.0 S4MIL 96.5 90.2 88.6 93.7 44.4 41.7 RRTMIL 94.8 89.3 87.3 92.8 49.7 44.3 MambaMIL 95.5 89.2 86.4 95.9 47.9 46.4 CAMIL 94.6 88.8 85.2 92.9 40.4 30.5 RPMIL 96.0 90.5 87.8 94.4 43.9 40.6 ACMIL 95.6 89.7 86.9 93.9 41.9 39.7 Jigsaw 96.3 90.7 88.0 94.7 44.3 40.9 LEAF-Mamba 96.9 91.3 88.4 95.3 48.9 45.2 Best scores in each column are in bold. |
| (c) Average Method AUC ACC F1 MeanMIL 74.2 45.2 35.4 MaxMIL 78.2 46.1 42.4 CLAM 84.4 52.1 47.8 ABMIL 83.3 50.9 45.7 MHIM-ABMIL 83.7 52.1 45.8 DSMIL 79.2 48.7 42.9 DTFD-MIL 82.8 48.8 44.5 TransMIL 81.1 47.8 40.0 MHIM-TransMIL 81.7 49.2 40.1 S4MIL 83.9 52.4 45.5 RRTMIL 83.6 53.5 46.3 MambaMIL 86.2 53.7 49.0 CAMIL 84.0 51.5 43.5 RPMIL 86.3 53.9 48.2 ACMIL 85.6 52.9 47.5 Jigsaw 86.6 54.4 48.5 LEAF-Mamba 87.7 56.1 50.0 Best scores in each column are in bold. Average summarizes the four diagnostic settings. |
Table 1: Diagnostic classification results with ResNet-50 features. All scores are percentages; higher is better.
| (a) BRACS and BRACS official BRACS BRACS official Method AUC ACC F1 AUC ACC F1 MeanMIL 73.2 43.3 33.7 61.2 32.0 23.2 MaxMIL 72.1 38.7 26.8 59.0 20.0 17.8 CLAM 75.9 40.7 31.0 65.6 31.0 22.9 ABMIL 76.9 42.5 34.4 69.1 30.0 26.8 MHIM-ABMIL 77.6 43.8 35.8 69.8 30.0 26.9 DSMIL 71.4 33.1 25.1 61.2 17.0 14.4 DTFD-MIL 77.1 42.5 34.7 70.1 26.0 23.8 TransMIL 71.2 35.3 24.0 67.1 29.0 25.1 MHIM-TransMIL 72.5 37.5 26.5 68.2 29.0 25.8 S4MIL 75.2 43.3 33.1 64.3 26.0 19.7 RRTMIL 78.5 45.5 35.5 71.5 28.0 24.2 MambaMIL 70.5 33.0 23.8 72.8 29.0 27.3 CAMIL 77.8 43.8 34.1 70.2 30.0 25.0 RPMIL 80.4 47.2 36.7 74.2 35.0 30.0 ACMIL 79.8 46.5 36.3 73.6 34.0 29.7 Jigsaw 80.5 47.8 37.0 74.6 35.0 30.2 LEAF-Mamba 82.2 49.6 37.9 77.8 36.0 29.8 Best scores in each column are in bold. |
| (b) Camelyon16 and EBRAINS Camelyon16 EBRAINS Method AUC ACC F1 AUC ACC F1 MeanMIL 78.7 71.9 65.4 94.9 51.5 50.3 MaxMIL 77.1 71.0 67.3 92.3 43.9 42.1 CLAM 95.6 90.5 88.3 95.2 53.9 53.2 ABMIL 96.3 91.7 89.6 95.4 54.1 53.4 MHIM-ABMIL 96.6 90.5 89.7 95.8 54.3 53.2 DSMIL 80.5 73.9 67.3 93.3 47.7 46.3 DTFD-MIL 95.3 90.0 87.4 94.0 51.1 50.1 TransMIL 94.5 88.5 86.4 92.0 49.9 50.6 MHIM-TransMIL 95.1 89.5 86.9 92.8 50.2 50.4 S4MIL 96.6 92.2 90.3 93.2 50.9 49.5 RRTMIL 96.8 92.1 90.1 94.1 50.7 50.3 MambaMIL 95.2 88.2 86.2 94.5 51.3 50.5 CAMIL 96.0 91.4 89.1 92.8 48.6 45.0 RPMIL 97.0 92.8 90.9 94.4 50.0 50.4 ACMIL 96.7 92.6 90.7 93.9 49.1 50.5 Jigsaw 97.1 92.9 91.0 94.2 51.2 51.0 LEAF-Mamba 97.3 92.8 91.5 96.8 55.0 53.4 Best scores in each column are in bold. |
| (c) Average Method AUC ACC F1 MeanMIL 77.0 49.7 43.1 MaxMIL 75.1 43.4 38.5 CLAM 83.1 54.0 48.8 ABMIL 84.4 54.6 51.0 MHIM-ABMIL 85.0 54.7 51.4 DSMIL 76.6 42.9 38.3 DTFD-MIL 84.1 52.4 49.0 TransMIL 81.2 50.7 46.5 MHIM-TransMIL 82.2 51.6 47.4 S4MIL 82.3 53.1 48.1 RRTMIL 85.2 54.1 50.0 MambaMIL 83.3 50.4 47.0 CAMIL 84.2 53.5 48.3 RPMIL 86.5 56.3 52.0 ACMIL 86.0 55.6 51.8 Jigsaw 86.6 56.7 52.3 LEAF-Mamba 88.5 58.4 53.1 Best scores in each column are in bold. Average summarizes the four diagnostic settings. |
Table 2: Diagnostic classification results with PLIP features. All scores are percentages; higher is better.
| (a) BRACS and BRACS official BRACS BRACS official Method AUC ACC F1 AUC ACC F1 MeanMIL 78.6 48.2 38.4 62.8 33.0 19.1 MaxMIL 79.5 47.8 36.4 69.2 33.0 26.3 CLAM 79.8 49.6 36.3 74.0 37.0 25.3 ABMIL 77.8 50.6 40.4 75.6 37.0 27.4 MHIM-ABMIL 78.4 51.2 41.5 76.1 37.0 28.7 DSMIL 78.7 46.7 35.6 74.9 37.0 29.9 DTFD-MIL 80.9 49.2 39.2 76.3 35.0 27.4 TransMIL 73.1 42.1 24.8 71.4 37.0 27.6 MHIM-TransMIL 74.5 44.5 28.5 72.5 37.0 28.5 S4MIL 77.4 46.5 36.2 72.5 33.0 22.5 RRTMIL 79.5 48.5 37.9 76.8 34.0 22.5 MambaMIL 78.0 47.0 32.6 76.6 35.0 27.7 CAMIL 79.7 49.8 37.5 75.2 37.0 27.0 RPMIL 81.7 51.6 42.5 77.5 38.0 31.4 ACMIL 81.2 51.0 41.6 77.2 38.0 30.7 Jigsaw 82.0 52.0 43.0 77.6 38.0 31.8 LEAF-Mamba 83.2 53.6 44.9 78.6 40.0 32.7 Best scores in each column are in bold. |
| (b) Camelyon16 and EBRAINS Camelyon16 EBRAINS Method AUC ACC F1 AUC ACC F1 MeanMIL 91.5 93.9 88.3 97.2 63.6 62.6 MaxMIL 87.7 90.2 89.7 97.4 67.6 67.0 CLAM 94.8 92.7 92.3 97.8 66.9 65.9 ABMIL 96.5 93.2 91.5 97.7 67.3 66.4 MHIM-ABMIL 96.9 93.4 91.5 97.9 67.9 67.1 DSMIL 92.7 88.5 84.8 97.8 66.7 65.9 DTFD-MIL 97.5 94.2 92.6 97.6 67.9 67.0 TransMIL 96.5 90.5 88.8 97.3 60.4 58.4 MHIM-TransMIL 96.9 91.7 89.7 97.6 61.8 59.7 S4MIL 95.8 94.0 92.5 97.5 63.2 62.5 RRTMIL 96.2 92.5 92.1 97.2 63.8 62.9 MambaMIL 97.0 92.5 90.7 97.2 64.3 62.6 CAMIL 96.2 92.9 91.8 96.1 62.7 64.5 RPMIL 97.8 94.0 92.4 95.8 64.2 65.7 ACMIL 97.5 93.8 92.2 96.1 64.4 64.3 Jigsaw 97.8 94.0 92.5 96.2 64.8 66.3 LEAF-Mamba 97.9 94.2 92.5 97.9 68.2 68.5 Best scores in each column are in bold. |
| (c) Average Method AUC ACC F1 MeanMIL 82.5 59.7 52.1 MaxMIL 83.4 59.7 54.9 CLAM 86.6 61.6 55.0 ABMIL 86.9 62.0 56.4 MHIM-ABMIL 87.3 62.4 57.2 DSMIL 86.0 59.7 54.1 DTFD-MIL 88.1 61.6 56.6 TransMIL 84.5 57.5 49.9 MHIM-TransMIL 85.4 58.8 51.6 S4MIL 85.8 59.2 53.4 RRTMIL 87.4 59.7 53.9 MambaMIL 87.2 59.7 53.4 CAMIL 86.8 60.6 55.2 RPMIL 88.2 62.0 58.0 ACMIL 88.0 61.8 57.2 Jigsaw 88.4 62.2 58.4 LEAF-Mamba 89.4 64.0 59.6 Best scores in each column are in bold. Average summarizes the four diagnostic settings. |
Table 3: Diagnostic classification results with CONCH v1.5 features. All scores are percentages; higher is better.
| Five-cohort average Method ResNet-50 PLIP CONCH v1.5 CLAM 58.8 59.0 64.4 ABMIL 59.1 60.9 62.5 TransMIL 65.5 64.9 63.9 S4MIL 65.1 63.0 64.7 RRTMIL 66.0 63.7 65.2 MambaMIL 66.7 66.0 65.5 CAMIL 65.7 65.2 64.6 RPMIL 66.1 66.2 65.6 ACMIL 66.0 65.7 65.0 Jigsaw 66.9 67.1 67.0 LEAF-Mamba 67.0 67.3 67.3 Best scores in each column are in bold. |
Table 4: Survival prediction results averaged over five TCGA cohorts. Values are C-index ; higher is better.
| Branch and fusion ablation ID LEA Branches Fusion AUC ACC F1 F1 Yes Local only – 75.7 35.3 27.4 F2 Yes Global only – 76.8 36.4 29.2 F3 Yes Global + local Concat 77.2 37.0 30.4 F4 Yes Global + local Add 77.0 36.9 31.0 F5 Yes Global + local Mean 77.1 36.8 30.8 F6 Yes Global + local Equal gate 77.5 37.6 31.4 F7 Yes Global + local Static gate 77.8 37.9 31.3 F8 Yes Global + local Gated 78.6 40.0 32.7 F9 – Global + local Concat 76.7 37.0 29.8 LEA denotes latent evidence adaptation. F8 is the complete model. Best scores in each metric are in bold. |
Table 5: Branch and fusion comparisons on BRACS official with CONCH v1.5 features. Values are percentages; higher is better.
| Adapter performance and parameter counts Parameters (K) Adapter Train Infer. AUC ACC F1 No adapter (frozen PCA bridge) 0.0 0.0 75.1 34.4 27.4 Linear projection 196.9 196.9 75.5 36.4 27.9 Two-layer MLP 721.9 721.9 76.8 36.1 28.3 Residual bottleneck adapter 722.4 722.4 76.4 35.7 28.5 LEA without reconstruction 723.5 723.5 75.8 35.8 28.1 Full LEA 1250.0 723.5 76.8 36.4 29.2 K denotes 1,000 parameters; counts are rounded to one decimal place. Training includes the reconstruction decoder; inference excludes it. Best performance scores are in bold. |
Table 6: Adapter comparisons on BRACS official with CONCH v1.5 features under global-only aggregation. Performance values are percentages.
| (a) BRACS official / CONCH v1.5 ID LEA Aggregation AUC ACC F1 C1 – Global only 76.6 35.3 27.7 C2 Yes Global only 76.8 36.4 29.2 C3 – Global + local ∗ 77.2 37.4 31.6 C4 Yes Global + local ∗ 78.6 40.0 32.7 |
| (b) BRACS official / PLIP ID LEA Aggregation AUC ACC F1 C1 – Global only 76.9 35.9 26.5 C2 Yes Global only 75.5 34.4 27.6 C3 – Global + local ∗ 77.2 29.6 29.3 C4 Yes Global + local ∗ 77.8 36.0 29.8 |
| (c) Camelyon16 / CONCH v1.5 ID LEA Aggregation AUC ACC F1 C1 – Global only 97.2 92.2 89.9 C2 Yes Global only 97.8 92.7 89.7 C3 – Global + local ∗ 97.5 93.8 92.7 C4 Yes Global + local ∗ 97.9 94.2 92.5 |
| (d) TCGA-BRCA / CONCH v1.5 ID LEA Aggregation C-index C1 – Global only 67.2 C2 Yes Global only 68.1 C3 – Global + local ∗ 66.0 C4 Yes Global + local ∗ 68.0 ∗ denotes gated global–local fusion. C4 is the complete model. Best scores in each setting are in bold. |
Table 7: Cross-setting component comparisons. Classification values are percentages; survival values are C-index . Higher is better.
Figure 5: Qualitative attention comparison on a BRACS malignant carcinoma case. The displayed top- patches differ between the models, and LEAF-Mamba predicts the correct malignant carcinoma label in this example.
More in this topic
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