Selective data curation enables efficient pretraining of chest radiograph foundation models
Year: 2026
Type: preprint
Abstract: Foundation models for medical imaging are pretrained on increasingly large datasets following a “scale-at-all-costs” paradigm, although medical corpora often contain substantial redundancy, class imbalance and heterogeneous data quality. Here we investigate whether selective data curation can reduce the data and computational requirements of chest X-ray (CXR) vision-language pretraining. We introduce CheXficient, which leverages representation-guided data curation to prioritize informative and important image-report pairs. CheXficient is pretrained using 22.7% of 1,235,004 paired CXR images and reports and under 27.3% of the total compute budget, yet maintains strong performance relative to full-data pretraining and remains competitive with other large-scale pretrained models. We assess CheXficient across 20 individual benchmarks spanning 5 task types, including non-adapted off-the-shelf evaluations (zero-shot findings classification and cross-modal retrieval) and adapted downstream tasks (disease diagnosis, semantic segmentation, and radiology report generation). Further analyses show CheXficient systematically prioritizes under-represented training samples, improving generalizability on long-tailed or rare thoracic conditions. These findings support selective data curation as a practical strategy for resource-efficient development of medical foundation models.
Source: Research Square
Institutions: Stanford University, Stanford Medicine
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7208746920