Pan-Organ Tumor Detection on Whole Slide Images

About the Research

This was a prestigious collaborative research project between IIT Bombay Medical Lab and the Tata Memorial Cancer Center. The goal was to develop a robust, generalizable computer vision model capable of detecting tumors across multiple distinct human organs using high-resolution, gigapixel Whole Slide Images (WSIs).

My Role & Contributions

As a Machine Learning Researcher, I led the model architecture design and the complex data preprocessing pipelines required for handling massive WSI files. I worked closely with medical professionals to ensure the output was clinically interpretable.

Technical Challenges Solved

• Fine-tuned KimiaNet (a histopathology-pretrained CNN) utilizing a DenseNet-121 backbone, achieving an 89% AUC on the multi-organ test set.
• Engineered an adaptive patch extraction pipeline capable of parsing gigapixel WSIs efficiently without exhausting VRAM.
• Implemented Reinhard color normalization to account for varying stain intensities across different hospital labs, ensuring cross-organ generalization (kidney, breast, colon).

Technologies Used

Lessons Learned

In medical AI, data preprocessing (like color normalization and artifact removal) often plays a larger role in model generalization than the raw architecture itself. Working with the Tata Memorial Center highlighted the critical need for model explainability when dealing with clinical diagnoses.