Deep learning quantifies pathologists’ visual patterns for whole slide image diagnosis
Problem
Training deep learning diagnostic models on whole slide images requires massive pixelwise annotation time from pathologists, limiting dataset size and diagnostic precision.
Why it matters
Pathologist annotation burden constrains the scale and accuracy of AI cancer diagnostics.
What the AI did
A deep learning system (PEAN) decodes pathologists' expertise from eye-tracking visual patterns and uses it to diagnose whole slide images with minimal manual annotation.
Results
Eye-tracking reduced WSI annotation time to 4% of manual annotation; evaluated on 5881 WSIs across 5 skin lesion categories with high AUC (value truncated in abstract).
Impact
Could sharply cut pathologist annotation workload while enabling precise whole-slide diagnostics.
Record
| Venue | Nature Communications |
|---|---|
| Published | 2025 |
| Work type | article |
| Authors | Tianhang Nan, Song Zheng, Siyuan Qiao, Hao Quan, Xin Gao, Xin Gao, Jun Niu, Bin Zheng et al. |
| Citations | 45 (OpenAlex, at import time) |
| References | 52 |
| Model or method | PEAN (Pathology Expertise Acquisition Network), deep learning |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1038/s41467-025-60307-1 |
| Open access | gold |
| Keywords | deep learning, whole slide images, eye-tracking, pathology, skin lesions, diagnostics |
Importance rating
significant — thousands to millions affected
Cuts pathologist WSI annotation time to 4%, easing a stated manpower bottleneck; evaluated on 5881 WSIs with 0.992 AUC, aiding large-scale data collection and clinical care.