AI & Machine Learning Overview
Lab 4 of the System Engineering track (event day 11: “AI — Data Processing, ML Modeling”) covers data processing and ML modelling with Python and PyTorch. The lab is taught end to end through the Kaggle RSNA Screening Mammography Breast Cancer Detection competition: roughly 54,000 DICOM mammograms from about 11,900 patients, a heavily imbalanced target (~2% positive), and a probabilistic F1 (pF1) metric that scores predicted probabilities rather than hard thresholds.
You will move from raw medical images to a cross-validated classifier, then upgrade the preprocessing stage the way the competition’s winning team did.
What you’ll build
The lab is split across two hands-on parts.
- Part 1: Mammography Baseline — DICOM decoding and normalisation, threshold-based breast ROI cropping, mammography-specific augmentations, a ConvNeXt-small classifier, and 4-fold patient-stratified cross-validation scored with pF1.
- Part 2: YOLOX ROI Pipeline — the winning-pipeline upgrade: train a YOLOX-nano detector to find the breast bounding box, produce cleaner and more consistent crops, then retrain ConvNeXt and compare against the baseline.
Both parts share a single workbook. Each section ends with TODO exercises that ask you to extend, probe, or measure the pipeline rather than just run it.
Get the notebook
The complete hands-on workbook (Parts 1 & 2) with TODO exercises. Run on Kaggle or a GPU workstation.
The dataset comes from the Kaggle RSNA Screening Mammography Breast Cancer Detection competition. You need a Kaggle account and must accept the competition rules before you can download it. The images are large (2048×1024) and training is memory-intensive — a GPU is strongly recommended.