Research map / Image generation / Autoregressive and latent generation

Fast and Accurate Gigapixel Pathological Image Classification with Hierarchical Distillation Multi-Instance Learning

CVPR 2025 · Image generation › Autoregressive and latent generation · approach: Retrieval and matching with generative features · 11 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 HDMIL, a hierarchical distillation MIL framework where a low-resolution pre-screening network (LIPN) distilled from a high-resolution dynamic MIL network (DMIN) skips irrelevant high-resolution patches, plus a Chebyshev-polynomial Kolmogorov-Arnold classifier, for fast and accurate WSI classification.

Problem

Multi-instance learning for gigapixel WSI classification has high inference cost because all high-resolution patches must be cropped and feature-extracted, and existing MIL methods cannot discard irrelevant patches before computing attention scores.

Approach

Use self-distillation between teacher (all instances) and student (selected instances) branches in DMIN to generate binary relevance masks, distill these masks into a lightweight low-resolution pre-screener LIPN that decides which high-resolution patches to skip at inference, and replace linear classifiers with a Chebyshev-polynomial KAN classifier.

Key results

• On Camelyon16, HDMIL reaches 90.88% AUC and 88.61% accuracy, 3.13% AUC above the previous best, while cutting inference time by 28.6%.
• HDMIL† (DMIN only) reaches 93.17% AUC on Camelyon16 test, the best AUC in that comparison, and HDMIL reduces inference time by 21.8% on TCGA-NSCLC and 7.2% on TCGA-BRCA relative to HDMIL†.
• Ablation on Camelyon16 validation shows the CKA classifier and self-distillation raise average AUC from 92.95 to 95.60 when combined with LIPN.

Measured effects reported in the paper

Surprising findings

• Discarding irrelevant instances improved classification performance rather than trading off against speed, as claimed by the authors.
• Performance on Camelyon16 test dropped for HDMIL relative to HDMIL† (90.88 vs 93.17 AUC), which the authors attribute to validation-set selection bias on a small dataset.

Limitations

• Test-set AUC drops on Camelyon16 when pre-screening is used, attributed to validation-set selection bias on a small dataset.
• Feature extractor (ResNet-50) cost remains the bottleneck and is not reduced by the method.
• Retention ratio r must be preset; performance peaks at an intermediate value and inference time rises with r.

Experimental setup

Datasets: Camelyon16 TCGA-NSCLC TCGA-BRCA

Benchmarks: Camelyon16 TCGA-NSCLC TCGA-BRCA

Metrics: AUC accuracy inference time (s)

Baselines: Max-Pooling Mean-Pooling ABMIL CLAM DSMIL TransMIL DTFD-MIL S4MIL MambaMIL SMT

Builds on: CLAM DynamicViT KAN ABMIL MobileNetV4

Closest papers by meaning

• CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.88)
• PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.87)
• Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.87)
• SI-MIL: Taming Deep MIL for Self-Interpretability in Gigapixel Histopathology (CVPR 2024 · similarity 0.86)
• FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.86)
• ViLa-MIL: Dual-scale Vision-Language Multiple Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.86)