Research map / Architectures and efficiency / Attention and transformers
Rotation-Agnostic Image Representation Learning for Digital Pathology
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 FPS, a KDE-based fast patch selection method, PathDino, a 5-block ~9M-parameter histopathology ViT, and HistoRotate, a 360° rotation augmentation within DINO self-supervised pretraining on ~6M TCGA patches.
Problem
Histopathology WSI analysis suffers from high computational cost of brute-force patching and MIL, and generic or large ViT feature extractors overfit and ignore histopathology-specific characteristics such as rotation invariance.
Approach
Combines density-proportional KDE patch selection with spatial constraints, a compact 5-block ViT, and a 360° rotation augmentation (HistoRotate) inside DINO self-supervised training so that rotated histopathology patches keep their context and improve embeddings.
Key results
- PathDino-512 reaches 83% WSI top-1 accuracy on Private-Liver and 58% on PANDA, versus 68% and 32% for HIPT.
- PathDino-512 achieves 88.57±3.08 macro-F1 on Private-Breast 5-fold cross-validation, above iBOT-Path's 85.12±1.74.
- FPS reduces patching time on Private-Breast from 20.4 to 13.1 minutes and on PANDA cuts missed WSIs from 268 to 138 compared with Yottixel.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PathDino-512 instead of iBOT-Path | PathDino-512 versus iBOT-Path · Private-Breast | accuracy (Top-1): 0.64 -> 0.68 | Helped |
| Use PathDino-512 instead of iBOT-Path | PathDino-512 versus iBOT-Path · Private-Liver | accuracy (Top-1): 0.79 -> 0.83 | Helped |
| Use PathDino-512 instead of DinoSSLPathology-8 | PathDino-512 versus DinoSSLPathology-8 · Private-Skin | accuracy (Top-1): 0.78 -> 0.78 | No clear effect |
| Use PathDino-512 instead of KimiaNet | PathDino-512 versus KimiaNet · PANDA | accuracy (Top-1): 0.57 -> 0.58 | Helped |
| Use PathDino-512 instead of KimiaNet | PathDino-512 versus KimiaNet · CAMELYON16 | accuracy (Top-1): 0.76 -> 0.73 | Hurt |
| Use PathDino-512 instead of iBOT-Path | PathDino-512 versus iBOT-Path · BRACS | accuracy (Top-1): 0.64 -> 0.64 | No clear effect |
| Use FPS instead of Yottixel mosaic with PathDino | PathDino-backed WSI patch selection · Private-Breast | accuracy (Top-1): 63 -> 68 | Helped |
| Use FPS instead of Yottixel mosaic with PathDino | PathDino-backed WSI patch selection · Private-Liver | accuracy (Top-1): 81 -> 83 | Helped |
| Use FPS instead of Yottixel mosaic with PathDino | PathDino-backed WSI patch selection · Private-Skin | accuracy (Top-1): 79 -> 78 | Hurt |
| Use FPS instead of Yottixel mosaic with PathDino | PathDino-backed WSI patch selection · Private-CRC | accuracy (Top-1): 57 -> 63 | Helped |
| Use FPS instead of Yottixel mosaic with PLIP | PLIP-backed WSI patch selection · PANDA | accuracy (Top-1): 53 -> 56 | Helped |
| Use FPS instead of Yottixel mosaic with PathDino | PathDino-backed WSI patch selection · CAMELYON16 | accuracy (Top-1): 76 -> 73 | Hurt |
Surprising findings
- On Private-Skin, iBOT-Path (73.09 macro-F1) outperforms PathDino-512 in 5-fold cross-validation.
- FPS is not uniformly better than Yottixel: on some datasets (e.g., Private-Liver Top-1 with PathDino) its accuracy is lower, and public-dataset results are described as on par.
Limitations
- Training data includes only 20x magnification patches from TCGA; multi-resolution training not explored.
- Public-dataset FPS results are on par rather than superior to Yottixel.
- iBOT-Path outperforms PathDino on Private-Skin in 5-fold cross-validation.
Experimental setup
Datasets: TCGA PANDA CAMELYON16 BRACS DigestPath PanNuke Kather-7K WSSS4LUAD
Benchmarks: PANDA CAMELYON16 BRACS DigestPath Kather PanNuke WSSS4LUAD
Metrics: top-1 accuracy majority vote MV@3 majority vote MV@5 macro F1 accuracy
Baselines: Yottixel mosaic HIPT iBOT-Path DinoSSLPathology PLIP BiomedCLIP KimiaNet DinoV2-ViT-b14
Builds on: DINO Yottixel iBOT HIPT
Closest papers by meaning
- Sequential Attention-based Sampling for Histopathological Analysis (NeurIPS 2025 · similarity 0.88)
- Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning (CVPR 2025 · similarity 0.88)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.88)
- Dynamic Entity-Masked Graph Diffusion Model for Histopathology Image Representation Learning (AAAI 2025 · similarity 0.86)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.86)
- Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.