Repository for SydneyMTL: Interpretable Multi-Task Learning for Complete Sydney System Assessment in Gastric Biopsies
- Unified multi-task learning for the Updated Sydney System (USS): A single weakly supervised MIL framework predicts all five USS attributes simultaneously: H. pylori, Intestinal Metaplasia, Glandular Atrophy, Neutrophil Activity, and Mononuclear Cell Infiltration, following the 4-tier severity grading (0–3). Atrophy additionally includes an explicit N/A class to reflect real-world diagnostic workflow.
- Task-specific attention for interpretability: The model uses shared slide representations with task-specific attention pooling and classification heads, enabling attribute-wise heatmaps that highlight regions contributing to each grading decision.
- Long-tail aware optimization via prior-based logit adjustment: To address severe class imbalance (dominance of Absent/Mild cases), we incorporate empirical class priors directly into logits. This improves robustness on rare Moderate/Marked grades and maintains performance under balanced evaluation.
- Emergent ordinal structure in representation space: Although trained with standard classification loss, the learned embeddings preserve biological ordinality—severity grades form a continuous trajectory in latent space.
- Large-scale validation: Evaluated on 50,765 retrospective WSIs and a 366-case expert-consensus “Golden” dataset with balanced severity distribution.
- Robust performance under balanced evaluation: While baseline methods degrade substantially on the Golden set, our model maintains high agreement (e.g., QWK up to 0.898 for IM and 0.826 for H. pylori).
- Clinically meaningful agreement with pathologists: Across 24 pathologists, the model achieves strong concordance (mean QWK ≈ 0.73), reflecting real-world diagnostic consistency.
- AI-assisted reading improves consistency and efficiency: In a randomized reader study, AI support increased inter-observer agreement and reduced reading time by ~34% per WSI.
- Pathologically plausible explanations: Attention maps localize biologically meaningful structures (e.g., goblet cells for IM, neutrophil infiltration for activity), supporting clinical interpretability.
| Name | ORCID | Affiliation | Notes | |
|---|---|---|---|---|
| Ho Heon Kim | 0000-0001-7260-7504 | hoheon0509@mf.seegene.com |
|
Contributed equally |
| Won Chang Jeong | 0009-0008-1931-5957 | jeongwonchan53@gmail.com |
|
Contributed equally |
| Yuri Hwang | - | - |
|
|
| Gisu Hwang | 0000-0003-1046-9286 | gshwang@mf.seegene.com |
|
|
| Kyungeun Kim | - | kekim@mf.seegene.com |
|
Corresponding author |
| Young Sin Ko | 0000-0003-1319-4847 | noteasy@mf.seegene.com |
|
Corresponding author |
-
$^{1}$ AI Research Center, Seegene Medical Foundation, 288 Dapsimni-ro, Seoul, South Korea -
$^{2}$ Pathology Center, Seegene Medical Foundation, 288 Dapsimni-ro, Seoul, South Korea