AC-MIL

Main results

eGFR prediction performance

AC-MIL with ABMIL achieves the highest Pearson and Spearman correlations and the lowest RMSE among the 21 evaluated configurations. The comparison includes whole-slide pooling, compartment-specific models, late fusion, and unidirectional gating.

Three-fold patient-level cross-validation across three seeds. Values reproduce the manuscript’s reported means and ± terms; bold marks the best value in each column. Slides from the same patient remain in the same fold.

AC-MIL reduces RMSE from 21.70 to 19.07 relative to non-glomerular ABMIL, a 12.1% reduction. Relative to whole-slide ABMIL, RMSE decreases by 12.9% (21.90 to 19.07), while Pearson r increases from 0.5214 to 0.6236.

Full comparison: all 21 configurations

Input / integration	Representation	Pooling	Pearson r ↑	Spearman ρ ↑	RMSE ↓
Bidirectional gate	Dual-Stream	TransMIL	0.5048 ± 0.0235	0.4684 ± 0.0413	22.53 ± 3.4903
AC-MIL	Dual-Stream	ABMIL	0.6236 ± 0.0506	0.5497 ± 0.0748	19.07 ± 3.3059
Bidirectional gate	Dual-Stream	Mean Pool	0.5248 ± 0.0477	0.4893 ± 0.0793	22.14 ± 2.8999
Glomerular gate	Dual-Stream	TransMIL	0.4612 ± 0.0410	0.4165 ± 0.0378	23.24 ± 3.6077
Glomerular gate	Dual-Stream	ABMIL	0.4978 ± 0.0280	0.4521 ± 0.0452	22.31 ± 2.6821
Glomerular gate	Dual-Stream	Mean Pool	0.5083 ± 0.1149	0.4839 ± 0.0975	22.30 ± 4.3130
Non-glomerular gate	Dual-Stream	TransMIL	0.4964 ± 0.0187	0.4815 ± 0.0328	23.28 ± 2.2006
Non-glomerular gate	Dual-Stream	ABMIL	0.4716 ± 0.0421	0.4493 ± 0.0525	22.88 ± 2.6494
Non-glomerular gate	Dual-Stream	Mean Pool	0.4624 ± 0.0281	0.4174 ± 0.0430	22.97 ± 2.5062
Late fusion	Dual-Stream	TransMIL	0.4939 ± 0.0271	0.4498 ± 0.0663	22.99 ± 3.0887
Late fusion	Dual-Stream	ABMIL	0.4938 ± 0.0088	0.4767 ± 0.0115	23.65 ± 2.8058
Late fusion	Dual-Stream	Mean Pool	0.4669 ± 0.0682	0.4321 ± 0.0841	22.58 ± 3.3420
Whole slide	Single-Stream	TransMIL	0.4734 ± 0.0073	0.4311 ± 0.0108	23.81 ± 2.3466
Whole slide	Single-Stream	ABMIL	0.5214 ± 0.0400	0.4850 ± 0.0434	21.90 ± 3.0088
Whole slide	Single-Stream	Mean Pool	0.4993 ± 0.0625	0.4671 ± 0.0781	22.47 ± 3.4559
Glomerular-enriched	Single-Stream	TransMIL	0.4805 ± 0.0585	0.4337 ± 0.0566	22.50 ± 4.1410
Glomerular-enriched	Single-Stream	ABMIL	0.4380 ± 0.0729	0.4001 ± 0.0780	23.30 ± 3.5793
Glomerular-enriched	Single-Stream	Mean Pool	0.4874 ± 0.0143	0.4528 ± 0.0395	22.45 ± 2.8337
Non-glomerular	Single-Stream	TransMIL	0.5031 ± 0.0456	0.4729 ± 0.0518	22.51 ± 3.0477
Non-glomerular	Single-Stream	ABMIL	0.5478 ± 0.0579	0.5165 ± 0.0504	21.70 ± 3.1528
Non-glomerular	Single-Stream	Mean Pool	0.4802 ± 0.0946	0.4579 ± 0.0944	22.63 ± 4.1248

The comparison includes TransMIL, ABMIL, and mean pooling. AC-MIL with ABMIL has the highest Pearson and Spearman correlations and the lowest RMSE.

Disease-specific experiments

Performance by disease group

The best-performing single-stream compartment differs between disease groups: glomerular-enriched tissue in diabetic kidney disease, and non-glomerular tissue in acute tubular injury and acute interstitial nephritis. AC-MIL achieves the highest Pearson correlation in both groups.

All comparisons use ABMIL. Bold denotes the highest correlation or lowest RMSE within each group. Compartment-only models achieve the lowest subgroup RMSE; AC-MIL has the highest correlation and second-lowest RMSE in both.

Matched patch-budget analysis

All three anatomy-aware inputs reduce RMSE relative to whole-slide ABMIL at each tested fixed bag size. Non-glomerular input has the largest reduction at 16 and 32 patches; AC-MIL has the largest reduction at 64 patches.

Input	K = 16	K = 32	K = 64	Natural bags
AC-MIL	−0.4509	−0.5772	−0.5299	−2.83
Glomerular	−0.7073	−0.3259	−0.4075	+1.40
Non-glomerular	−0.7916	−0.6594	−0.5041	−0.20

Signed ΔRMSE relative to whole-slide ABMIL; more negative is better. Fixed-cardinality experiments use five seeds, with the same total patch budget and an even split across AC-MIL’s two streams. Non-glomerular input leads at K = 16 and K = 32; AC-MIL leads at K = 64.

Transferability

Transfer to proteinuria and albuminuria

Slide embeddings learned for eGFR prediction are frozen and used to train an MLP for proteinuria or albuminuria classification. These experiments assess whether the learned representations support related renal endpoints.

All single-stream representations use ABMIL; AC-MIL uses bidirectional gating. The eGFR-trained slide embeddings are frozen, and downstream MLP classifiers are trained for each target. Bold marks the best value per target and metric. This evaluates transfer between tasks; external-cohort generalization is not evaluated here.

Among single-stream representations, non-glomerular embeddings have the highest proteinuria AUROC, while glomerular embeddings have the highest albuminuria AUROC. AC-MIL has the highest AUROC and accuracy for both targets. Whole-slide embeddings have a slightly higher albuminuria F1 (0.7668 versus 0.7648).

Pathology analysis

Exploratory analysis of tissue morphology

High-attention patches from each stream are examined using a vision-language model, comparing patients with high and low predicted eGFR.

High and low refer to model-predicted eGFR. Quilt-LLaVA characterizes high-attention patches, and a separate language model aggregates the outputs. The reported uncertainty uses patch-level bootstrapping. Colored values mark the differences highlighted in the manuscript.

The VLM-based summaries identify different morphologic patterns across the two streams. These exploratory findings require validation by expert renal pathologists.

Citation

AC-MIL is presented in Context Matters: Anatomy-Aware Dual-Stream Multiple Instance Learning Framework for eGFR Prediction.

@inproceedings{arab2026context,
  title     = {Context Matters: Anatomy-Aware Dual-Stream Multiple Instance Learning Framework for e{GFR} Prediction},
  author    = {Abteen Arab and Maziar Riazy and Ali Bashashati},
  booktitle = {Medical Imaging with Deep Learning - Short Papers},
  year      = {2026},
  url       = {https://openreview.net/forum?id=KTfPFMntqQ}
}