Efficient X-Ray Representations For Classifying Diseases

UBC Medicine Datathon 2025 (Team 8)

Team Members:

• Ethan Rajkumar
• Joel Bonnie
• Pushya Jain
• Erhan Javed
• Vivaan Jhaveri
• Charity Grey

Chest X-rays, a primary diagnostic tool for identifying thoracic diseases, are the focus of our project, which aims to develop an efficient X-ray representation for automating and benchmarking weakly supervised classification and localization of common thoracic diseases. By leveraging clinical metadata, bounding box annotations, and efficient vector retrieval via Pinecone, we address challenges posed by large-scale datasets (~112k X-rays), annotation constraints, and weak supervision, and build robust classification models using Support Vector Machines (SVM) and Random Forest classifiers.

Overall metrics for positive classes:

• Overall positive recall: 0.1135
• Overall positive precision: 0.3384
• Overall positive F1 score: 0.1700
• Mean AUC ROC: 0.7242
• Mean Average Precision: 0.1253

---

Conda Environments:

Feature Eng

Create the conda environment: conda env create -f environment.yml

Activate the conda environment: conda activate xray-analysis

SVC Processing

Create the conda environment: conda env create -f inference_env.yml

Activate the conda environment:

conda activate xray-svc