shrmoon/fetal-us-quality-gen

Cross-domain generalization study for fetal ultrasound plane classification using PyTorch, testing model performance on a held-out imaging device for low-resource clinical settings. # Fetal Ultrasound Plane Classification: A Cross-Domain Generalization Study A data pipeline and modeling project analyzing a 12,400-image medical imaging dataset, focused on building a reproducible pipeline to measure how well models generalize across different data sources — and how much of that generalization gap can be closed through better model design. ## Overview Most machine learning models are evaluated on data that looks like their training data. In practice, data pulled from a different source, device, or time period often behaves differently — a common failure mode in any production data system. This project builds a small, reproducible pipeline to measure that gap directly: structuring a dataset by source, engineering a deliberate train/test split along that source boundary (rather than a random split), and quantifying how much model performance drops as a result. ## Dataset - **Source:** FETAL_PLANES_DB (Burgos-Artizzu et al., 2020), a public dataset of 12,400 medical images from 1,792 subjects, collected across multiple imaging devices at two hospitals. zenodo.org - **Structure:** a CSV of metadata (`FETAL_PLANES_DB_data.csv`) joined against image files, with fields including `Image_name`, `Patient_num`, `Plane` (target label), `US_Machine` (data source), and `Operator`. - Raw data is not redistributed in this repository, in line with the source dataset's terms of use. Setup instructions: `data/README.md`. ## Data Pipeline Design The core engineering problem this project solves: build a pipeline that tests generalization across data sources, not just overall accuracy. - **Ingestion:** load and join image files against CSV metadata using pandas, validating expected columns and image paths. - **Split strategy:** rather than a random train/test split, the pipeline deliberately partitions data by `US_Machine` (data source) — holding one source out entirely from training. This produces two distinct evaluation sets from the sa …