A New Era in Computational Pathology
Advancing cancer research through a modular preprocessing framework (SlideLab) and a novel interpretability method (PAC).
The Challenge with Whole Slide Images (WSIs)
Data Inconsistency
Variations in H&E staining and scanning resolutions across institutions create significant biases, hindering the development of generalizable AI models.
Pervasive Artifacts
WSIs are often contaminated with artifacts like pen marks, tissue folds, and out-of-focus regions that can mislead AI models and corrupt results.
The "Black Box" Problem
The inability to understand why a model makes a specific prediction erodes clinical trust and creates a barrier to real-world adoption.
Solution 1: SlideLab - A Flexible Preprocessing Pipeline
SlideLab is a modular, high-performance framework designed to standardize WSI data by robustly handling artifacts and inconsistencies. Its flexible architecture allows researchers to build custom pipelines tailored to their specific needs.
Slide-Level Processing
Tissue Segmentation & Artifact Removal
Tile-Level Processing
Stain Normalization & Quality Control
Feature Extraction
(Optional) ResNet50, UNI, etc.
High-Quality Dataset
Ready for Downstream Analysis
SlideLab Outperforms Existing Tools
When benchmarked against five other popular preprocessing pipelines, SlideLab demonstrated superior performance in tissue segmentation, achieving the highest Dice Score against a gold-standard ground truth.
Project Impact & Achievements
Project Analysis 🧠
- Tackles Core Issues: Addresses the critical needs for standardized preprocessing and interpretable AI.
- Dual Contributions: Delivers both a practical tool (`SlideLab`) and a novel analytical method (`PAC`).
Key Achievements 🏆
- Novel Tool Development: Built a complete, functional software pipeline from the ground up.
- Superior Performance: Achieved the highest Dice score (0.741) in rigorous benchmarking against 5 competitors.
- Improved Model Outcomes: Showed that data processed with SlideLab leads to better downstream model performance.
Summary of Results 📊
- Quantifiably Better Data: `SlideLab` produces higher quality data, proven by superior metrics.
- Deeper Biological Insight: `PAC` successfully validated model reasoning in both MSI and B-ALL case studies, linking predictions to known biology.
Solution 2: Polarized Attentional Confidence (PAC)
PAC is a novel interpretability method that enhances trust in "black box" models by identifying the image regions that are both **important** to the model's decision and classified with **high confidence**.
Attention Score
(Importance)
Probability Score
(Confidence)
PAC Score
(Interpretable Insight)
PAC in Action: Real-World Case Studies
Case Study 1: Microsatellite Instability (MSI)
In colorectal cancer, PAC confirmed that the AI model correctly identified regions with high lymphocyte infiltration as indicative of MSI-High tumors, aligning perfectly with known clinical biology. This validates that the model learned meaningful biological features.
Result: Increased trust and biological validation.
Case Study 2: B-cell Acute Lymphoblastic Leukemia (B-ALL)
When applied to flow cytometry data, PAC revealed that the model distinguished B-ALL cells from their benign mimics by focusing on cells with high expression of key biomarkers like CD10 and CD34.