Research map / Label-efficient and robust learning / Domain adaptation and generalization

Samba: Severity-aware Recurrent Modeling for Cross-domain Medical Image Grading

NeurIPS 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Cross-domain text and graph transfer · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Samba, a Mamba-based grading model that scans image patches recurrently and recalibrates features with EM-estimated GMM severity bases, for domain-generalized medical image grading.

Problem

Disease grading from medical images must generalize from a source domain to unseen target domains, where cross-domain style variation disrupts identification of the most severe lesion and within-/cross-level feature clustering.

Approach

Encode image patches as a sequence with bidirectional selective state space (Mamba) layers to propagate severity information, then apply an EM-based state recalibration that models patch features as a GMM with learnable severity bases (updated by moving average) to compact the feature space and reduce domain shift.

Key results

• On fatigue fracture grading (Domain-1 to Domain-2), Samba reaches 76.2% ACC, 81.5% AUC and 45.8% F1 versus 52.7%, 70.4% and 28.7% for VMamba.
• On the DR benchmark, Samba averages 69.2% ACC and 40.9% F1 across six single-source settings, versus 65.1% ACC and 38.5% F1 for VMamba-ERM and 46.7% ACC for the best prior domain-generalized method GDRNet.
• On breast cancer grading, Samba raises average accuracy from 40.4% to 54.8% for the tiny backbone (VMamba-T) with domain-1 as source.

Measured effects reported in the paper

Surprising findings

• Using a moving average to update severity bases outperforms back-propagation updates, which the authors attribute to instability from EM iterations.

Limitations

• Under severe class imbalance in the source domain, the estimated GMM lesion distribution may not reflect the domain-agnostic lesion distribution.
• Results do not include error bars or statistical significance tests.

Experimental setup

Datasets: Cross-domain Fatigue Fracture Grading Benchmark Cross-domain Breast Cancer Grading Benchmark DeepDR Messidor IDRID APTOS FGADR RLDR DDR EyePACS CAMELYON17

Benchmarks: Fatigue Fracture Grading Benchmark Breast Cancer Grading Benchmark Diabetic Retinopathy Grading Benchmark CAMELYON17

Metrics: ACC AUC F1

Baselines: VMamba-ERM VMamba Vision Transformer LSTM UR-LSTM GDRNet DRGen MIL-ViT Swin-Transformer Mixup MixStyle GREEN

Builds on: Mamba VMamba Vision Mamba GDRNet Expectation-Maximization

Closest papers by meaning

• A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images (AAAI 2025 · similarity 0.87)
• Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation (CVPR 2024 · similarity 0.85)
• Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data (CVPR 2025 · similarity 0.84)
• Perturbating, Tuning, and Collaborating: Harnessing Vision Foundation Models for Single Domain Generalization on Medical Imaging (AAAI 2025 · similarity 0.84)
• Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.84)
• Why Invariance is Not Enough for Biomedical Domain Generalization and How to Fix It (NeurIPS 2026 · similarity 0.84)