BCA Detect Project

Breast cancer detection using CNNs and computer vision on cytology slide data from the GDC, in collaboration with NUS researchers.

python ml jupyter medical-imaging
View on GitHub →

Overview

A collaboration with SG Code Campus and NUS researchers to explore the use of convolutional neural networks for breast cancer detection. The project applies computer vision techniques to cytology slide images, working with real clinical data from the Genomic Data Commons (GDC) managed by the National Cancer Institute.

The aim is to support ongoing research into how machine learning can assist in medical cancer studies, starting with automated analysis of breast cancer tissue samples.

Dataset

The project uses data from the BRCA project hosted on the Genomic Data Commons (GDC). The raw data comes as zoomable whole-slide images (.svs format) which need to be converted to standard image formats (JPEG/PNG) before they can be fed into neural networks.

A custom data processing pipeline handles the conversion and tiling of these high-resolution pathology slides into model-ready inputs.

Approach

  • Whole-slide images (.svs) converted and tiled into standard image patches
  • CNN-based classification pipeline built in Jupyter notebooks
  • Models trained on labeled cytology data to distinguish cancerous vs. normal tissue
  • Iterative experimentation with different architectures and hyperparameters

Tech Stack

  • Language: Python
  • Environment: Jupyter Notebooks
  • Key Libraries: PyTorch/TensorFlow, OpenSlide (for .svs processing), NumPy, Matplotlib
  • Data Source: NCI Genomic Data Commons (GDC)

What I Learned

This project was my introduction to working with real clinical data. I learned that data preprocessing is often the hardest part — converting whole-slide .svs images into usable patches required understanding both the medical imaging domain and the tooling (OpenSlide). Collaborating with NUS researchers taught me how to communicate technical ML concepts to domain experts, and how important it is to validate your pipeline against ground truth before trusting any model output.