This project involves identifying areas of actual cancer in digital pathology images and detecting a specific type of cancer (ALK-tumor)

Deep_Learning_Approaches_for_Identifying_ALK_positive_Tumors_in_Pathology_Images.pdf

About the Project

This project aims to develop a model specifically designed to accurately detect ALK-positive tumors in digital pathology images, even with a limited dataset. To address data scarcity, the project will focus on extracting refined features using the PLIP model, which is pre-trained for pathology, rather than commonly used models like ResNet. This approach will enhance both the model's explainability and accuracy in identifying ALK-positive tumors.

Workflow

1. WSI Preprocessing and Patch Extraction (CLAM):
   • Whole Slide Images (WSI) are processed using the CLAM framework.
   • CLAM identifies and extracts patches (256x256 pixels each) from regions with the highest probability of containing cancer.
   • For each patch, CLAM provides a probability score indicating how likely it is to contain cancer.
2. Filtering Patches:
   • Only patches with a cancer probability of 80% or higher are selected.
   • These selected patches form the dataset for further analysis.
3. Feature Extraction (PLIP and Model Comparison):
   • Features are extracted from the selected patches using the PLIP model.
   • Various models (ResNet, MobileNet, CLIP, ViT) are compared to determine which extracts the most relevant features for pathology images.
   • The goal is to find features that best represent the cancer patches, given the small dataset size, requiring more precise feature extraction.
4. Binary Classification Model:
   • A simple neural network with around 3 layers is constructed for binary classification.
   • The model is trained to distinguish between ALK-positive and ALK-negative cancer types.
5. Final Output:
   • The classification model can now detect and differentiate between ALK-positive and ALK-negative cancer in the selected patches.
   • Visualization: To enhance interpretability, the patches are visualized on the original WSI image.
     • The patches are overlaid on the WSI, with color-coding (e.g., red for ALK-positive, blue for ALK-negative).
     • This allows for a clear, visual representation of the regions where cancer was detected, highlighting the different classifications across the slide.
     • This step can help in understanding the spatial distribution of cancer within the WSI.

Tech Stack

Python Pytorch MedicalAI Classification Object Detection Multiple Instance Learning Digital Pathology

Role

All research was conducted independently with the professor’s assitance.

Conducted By