Research map / Architectures and efficiency / Attention and transformers
BioX-CPath: Biologically-driven Explainable Diagnostics for Multistain IHC Computational 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 BioX-CPath, an explainable GNN for multistain IHC whole-slide image classification with a Stain-Aware Attention Pooling module and biologically grounded interpretability metrics, evaluated on RA and Sjogren's datasets.
Problem
Classifying unregistered, unannotated multistain IHC whole-slide image stacks in computational pathology while providing biologically interpretable explanations aligned with known disease mechanisms.
Approach
A hierarchical GAT-based patch graph combining feature-space and region-adjacency edges, with a Stain-Aware Attention Pooling (SAAP) module that computes stain-level attention weights and stain-aware patient embeddings, plus derived metrics (stain attention, entropy, stain-stain interaction scores, GNN heatmaps) for biological interpretation.
Key results
- On the RA dataset BioX-CPath reached 0.90 accuracy vs 0.86 for MUSTANG; on Sjogren's it reached 0.84 accuracy and 0.88 AUC, outperforming all baselines.
- Ablation on Sjogren's: adding random-walk positional encoding raised accuracy from 0.756 to 0.80 and SAAP raised it to 0.84.
- On RA, BioX-CPath did not beat MUSTANG in AUC (0.96 vs 0.96 tie) or average precision (0.98 vs 0.97 reported as best).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BioX-CPath instead of Patch-GCN | RA classification · RA | accuracy: 0.83 -> 0.90 | Helped |
| use BioX-CPath instead of MUSTANG | RA classification · RA | AUC: 0.96 -> 0.96 | No clear effect |
| use BioX-CPath instead of MUSTANG | RA classification · RA | average precision: 0.97 -> 0.98 | Helped |
| use BioX-CPath instead of MUSTANG | Sjogren classification · Sjogren | accuracy: 0.80 -> 0.84 | Helped |
| use BioX-CPath instead of MUSTANG | Sjogren classification · Sjogren | AUC: 0.85 -> 0.88 | Helped |
| use BioX-CPath instead of MUSTANG | Sjogren classification · Sjogren | average precision: 0.84 -> 0.86 | Helped |
| add random-walk positional encoding (ablation) | BioX-CPath · Sjogren | accuracy: 0.756 (0.059) -> 0.80 (0.038) | Helped |
| add SAAP (ablation) | BioX-CPath · Sjogren | accuracy: 0.756 (0.059) -> 0.84 (0.018) | Helped |
| add MHSA (ablation) | BioX-CPath · Sjogren | accuracy: 0.756 (0.059) -> 0.736 (0.049) | Hurt |
| add random-walk positional encoding (ablation) | BioX-CPath · RA | accuracy: 0.79 (0.018) -> 0.86 (0.018) | Helped |
| add SAAP (ablation) | BioX-CPath · RA | accuracy: 0.79 (0.018) -> 0.90 (0.019) | Helped |
| add MHSA (ablation) | BioX-CPath · RA | accuracy: 0.79 (0.018) -> 0.78 (0.025) | Hurt |
Surprising findings
- Adding multi-head self-attention slightly decreased accuracy on both datasets (e.g., 0.756 to 0.736 on Sjogren's, 0.79 to 0.78 on RA).
- BioX-CPath did not outperform MUSTANG in AUC on RA (0.96 vs 0.96).
Limitations
- Evaluated only on two autoimmune datasets with relatively small patient counts (153 and 93 patients).
- MHSA component slightly reduced performance while adding interpretability.
- Biological interpretation is qualitative alignment with known pathology rather than validated against ground truth biology.
Experimental setup
Datasets: Rheumatoid Arthritis multistain IHC dataset (607 WSIs, 153 patients) Sjogren's Disease multistain IHC dataset (347 WSIs, 93 patients)
Benchmarks: RA multistain IHC dataset Sjogren's multistain IHC dataset
Metrics: accuracy AUC average precision macro F1 precision recall
Baselines: ABMIL CLAM-SB DeepGraphConv PatchGCN TransMIL GTP MUSTANG
Builds on: SAGPool GAT MUSTANG UNI
Closest papers by meaning
- CPathAgent: An Agent-based Foundation Model for Interpretable High-Resolution Pathology Image Analysis Mimicking Pathologists' Diagnostic Logic (NeurIPS 2025 · similarity 0.85)
- OT-StainNet: Optimal Transport Driven Semantic Matching for Weakly Paired H&E-to-IHC Stain Transfer (AAAI 2025 · similarity 0.85)
- Revisiting End-to-End Learning with Slide-level Supervision in Computational Pathology (NeurIPS 2025 · similarity 0.85)
- Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis (CVPR 2024 · similarity 0.85)
- CPath-Omni: A Unified Multimodal Foundation Model for Patch and Whole Slide Image Analysis in Computational Pathology (CVPR 2025 · similarity 0.84)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.84)
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.