Business Context
MediScan AI is building a clinical decision-support tool for hospital radiology teams processing roughly 8,000 chest X-rays per day. The goal is to prioritize likely pneumonia cases for faster review, not to replace radiologists.
Dataset
You are given a labeled chest X-ray dataset collected from 5 hospitals over 24 months. Images are grayscale DICOM files converted to PNG for training. Labels were derived from radiologist reports and later adjudicated on a subset.
| Feature Group | Count | Examples |
|---|
| Image data | 1 | 1024x1024 frontal chest X-ray |
| Patient metadata | 5 | age, sex, portable_exam_flag, hospital_id, scanner_vendor |
| Study metadata | 4 | study_date, inpatient_flag, AP/PA view, image_quality_score |
- Size: 68K studies, 1 image per study, 9 tabular metadata fields
- Target: Binary classification — pneumonia present (1) vs absent (0)
- Class balance: 18% positive, 82% negative
- Missing data: 7% missing in image_quality_score, 3% missing in scanner_vendor, occasional corrupted image files
Success Criteria
A strong solution should achieve AUROC >= 0.90, AUPRC >= 0.70, and recall >= 0.85 at precision >= 0.60 on a hospital-held-out test set. The model should also provide probability outputs suitable for threshold tuning.
Constraints
- Inference latency must be <150 ms per image on a single T4 GPU
- The system must generalize across hospitals and scanner vendors
- Interpretability is required at the image level (e.g., saliency/Grad-CAM)
- Training budget is limited to a single multi-GPU run per experiment cycle
Deliverables
- Build a deep learning pipeline for binary image classification
- Define preprocessing and augmentation appropriate for medical imaging
- Explain how you will prevent patient- or hospital-level leakage
- Evaluate performance with clinically relevant metrics and threshold selection
- Propose a deployment-ready inference and monitoring plan
