💻 computer science

Selective data curation enables efficient pretraining of chest radiograph foundation models

The paper introduces CheXficient, a chest X-ray foundation model that leverages representation-guided selective data curation to achieve competitive performance across diverse clinical tasks using only a fraction of the data and computational resources required by traditional large-scale pretraining approaches.

Original authors: Curtis Langlotz, Chong Wang, Yabin Zhang, Yunhe Gao, Maya Varma, Clemence Mottez, Faidra Patsatzi, Fuying Wang, Jiaming Liu, Jin Long, Jean-Benoit Delbrouck, Sergios Gatidis, Akshay Chaudhari

Published 2026-09-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Curtis Langlotz, Chong Wang, Yabin Zhang, Yunhe Gao, Maya Varma, Clemence Mottez, Faidra Patsatzi, Fuying Wang, Jiaming Liu, Jin Long, Jean-Benoit Delbrouck, Sergios Gatidis, Akshay Chaudhari

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of medical imaging, computers are learning to see what doctors see. For years, the prevailing belief has been that to make these artificial intelligence systems smarter, you simply need to feed them more data. The logic seemed sound: the more chest X-rays and the more radiology reports a computer reads, the better it should become at spotting diseases. This approach, often called "scale at all costs," has led researchers to gather millions of images from hospitals around the globe. However, this strategy comes with a heavy price. Training these massive systems requires enormous amounts of electricity and time, often taking weeks on powerful supercomputers. Furthermore, medical data is rarely perfect; it is often full of duplicates, skewed toward common diseases, and missing the rare but critical conditions that doctors need to identify most urgently. The question facing the field is whether there is a smarter way to teach these machines without burning through resources or getting lost in a sea of redundant information.

A team of researchers at Stanford University has proposed a different path, one that focuses on quality over quantity. They introduced a new method called CheXefficient, which acts like a careful editor for the data used to train medical AI. Instead of blindly feeding the computer every single X-ray and report it can find, this system uses a "data curator" to select only the most useful examples. Imagine a library where a librarian decides which books to keep on the shelves based on how much new information they offer, rather than just stacking every book that arrives. In this case, the computer looks at the images and reports it has already seen and identifies which new examples are most different from what it knows. It prioritizes images that show rare or unusual conditions and down-weights those that are common and repetitive. By doing this, the system learns a more complete picture of human health with far fewer examples.

The researchers tested this idea using a massive collection of over 1.2 million paired chest X-rays and their corresponding radiology reports. They trained three different versions of their AI model to compare results. The first version, which they call CheXfull, was trained on the entire dataset of 1.2 million images. The second, CheXrandom, was trained on a random selection of 280,000 images, representing a typical reduction in data size. The third, CheXefficient, was trained on a carefully curated set of 280,000 images, chosen specifically because they offered the most new information and covered the widest range of medical conditions. The results were striking. CheXefficient, using only about 23 percent of the data and less than 28 percent of the computing power required by the full model, performed just as well as the massive model trained on all 1.2 million images. In many tests, it actually outperformed the model trained on the random selection of data, proving that the specific choice of data mattered far more than the sheer volume.

This approach did more than just save time and energy; it made the AI better at spotting the rare diseases that are often missed. In the full dataset, common conditions like normal lungs or minor fluid buildup dominate the numbers, while rare issues like specific types of fractures or unusual infections appear very few times. Because the data curator actively sought out these under-represented examples, the CheXefficient model learned to recognize them much more effectively. When tested on diseases that were rare in the training data, the curated model showed a significant ability to generalize, meaning it could apply what it learned to new, unseen patients with greater accuracy. This suggests that by intentionally balancing the diet of information the AI consumes, researchers can build systems that are not only cheaper to train but also more reliable for the difficult cases that matter most in a hospital.

The benefits extended beyond just identifying diseases. The researchers also tested how well these models could perform other complex tasks, such as describing the anatomy in an image, segmenting specific organs, or even writing a radiology report from scratch. In every category, the model trained on the curated data held its own against the massive full-data model. It required far fewer labeled examples to learn these new tasks, meaning doctors would need to spend less time manually correcting the AI's work. The study also found that the curated data naturally included a better mix of patients from different age groups and backgrounds, particularly capturing more images of children and older adults, groups that are often under-sampled in standard medical datasets. This indicates that the selection process naturally improved the fairness and diversity of the AI's knowledge without needing to manually enforce strict quotas.

The findings challenge the long-held assumption that bigger is always better in the world of medical artificial intelligence. The researchers demonstrated that a strategic approach to data selection can achieve the same high performance with a fraction of the resources. This is a crucial development for the future of medical AI, as it suggests that hospitals and research centers with limited budgets can still build powerful, state-of-the-art diagnostic tools. By focusing on the most informative examples and avoiding the noise of redundant data, the path to smarter medical AI becomes more efficient and accessible. The study concludes that the future of these foundation models lies not in hoarding every available image, but in curating a smarter, more representative collection that teaches the computer to see the full spectrum of human health.

Technical Summary: Selective Data Curation Enables Efficient Pretraining of Chest Radiograph Foundation Models

Problem Statement

Medical foundation models for imaging typically adhere to a "scale-at-all-costs" paradigm, relying on massive datasets to learn generalizable representations. However, this approach faces significant practical barriers in clinical settings:

Data Inefficiency: Medical corpora often contain substantial redundancy, severe class imbalance (where normal findings and common pathologies dominate), and heterogeneous data quality. Blindly training on all available data biases representation learning toward over-represented patterns.
Computational Burden: Exhaustive training on millions of samples incurs prohibitive computational costs (e.g., thousands of TPU-hours or weeks of GPU time), creating barriers for academic and clinical institutions with limited resources.
Long-Tail Generalization: Standard pretraining often fails to adequately represent rare but clinically important conditions, leading to biased models that struggle with under-represented thoracic diseases.

The central question addressed is whether strategic data curation can substantially reduce data and computational requirements while maintaining or improving the performance of chest X-ray (CXR) vision-language foundation models.

Methodology: CheXﬁcient

The authors introduce CheXﬁcient, a framework that leverages representation-guided data curation to prioritize informative and important image-report pairs during the early stages of pretraining.

Core Mechanism

CheXﬁcient operates within a Contrastive Language-Image Pretraining (CLIP) framework using a DINOv2 image encoder and a BioClinicalBERT text encoder. The curation process occurs during the early pretraining stage and involves the following steps:

Prototype-Driven Selection: The model maintains a set of evolving prototypes (prototypical centroids) that characterize the current multimodal representation space.
Distance-Based Prioritization: Training samples are assigned to their nearest prototype. The distance ( $d_i$ $d_{i}$ ) of a sample from its assigned prototype determines its priority:
- High Priority (Under-represented): Samples far from prototypes (in "tail" or "peripheral" regions) correspond to rare pathologies, atypical imaging appearances, or under-represented semantics. These are emphasized.
- Low Priority (Redundant): Samples close to prototypes (in "central" regions) represent common, redundant patterns. These are down-weighted and under-sampled.
- Noise Filtering: Samples in the "outlier" region (extremely far distances) are discarded as potential noise (e.g., corrupted images or mismatched pairs).
Adaptive Stratification: The selection uses a Gaussian-inspired adaptive stratification within each prototype cluster, partitioning samples into central, peripheral, tail, and outlier regions. A generalized power allocation rule distributes the sampling budget to ensure geometric coverage while avoiding near-duplicates via farthest point sampling.
Iterative Update: Prototypes are updated concurrently with model training using an exponential moving average (EMA) of selected sample embeddings, allowing the curation criteria to evolve as the model's representation space changes.

Experimental Setup

Pretraining Corpus: A large-scale database of 1,235,004 paired CXR images and reports from 13 public datasets (e.g., MIMIC-CXR, CheXpert, VinDr-CXR).
Baselines:
- CheXfull: Pretrained on the entire 1.235M dataset.
- CheXrandom: Pretrained on a random subset of the same size as the curated set.
- CheXﬁcient: Pretrained on a curated subset of 280,000 pairs (22.7% of the full data).
Evaluation Protocols:
- Non-adapted (Zero-shot): Findings classification (47 findings) and cross-modal retrieval.
- Adapted (Fine-tuning): Multi-disease diagnosis, semantic segmentation (anatomical/pathological), and radiology report generation.

Key Results

1. Data and Compute Efficiency

Resource Reduction: CheXﬁcient achieves strong performance using only 22.7% of the pretraining data and under 27.3% of the total compute budget (measured in H100 GPU-hours) compared to CheXfull.
Performance Parity: Despite the reduced resources, CheXﬁcient maintains performance comparable to CheXfull across 20 evaluation benchmarks. In many zero-shot classification tasks, it significantly outperforms CheXfull (e.g., on SIIM-PTX, $p=0.031$ ) and consistently outperforms CheXrandom.

2. Representation Quality and Long-Tail Generalization

Feature Distribution: Analysis of feature embeddings reveals that the curated subset systematically prioritizes low-density regions of the data manifold. Curated samples exhibit larger average k-nearest neighbor (kNN) distances compared to the full set, indicating an enrichment of under-represented patterns.
Rare Disease Handling: CheXﬁcient demonstrates improved zero-shot generalization for a broad range of pathologies (16 out of 26 tested in VinDr-CXR), including "Pulmonary fibrosis," "Pleural effusion," and "Aortic enlargement," compared to non-curated baselines.
Few-Shot Adaptability: In k-shot learning scenarios (1–5 labeled samples), CheXﬁcient adapts more effectively to long-tailed conditions (e.g., Calcification, Enlarged PA) than CheXfull, achieving higher median AUROC scores.

3. Downstream Task Performance

Label Efficiency: In fine-tuning for disease diagnosis, CheXﬁcient matches the performance of CheXrandom using only ~10% of the labeled training data.
Segmentation: For semantic segmentation (pneumothorax, lungs, ribs), CheXﬁcient matches CheXfull performance while using only 10–30% of the densely annotated training samples.
Report Generation: In radiology report generation tasks, CheXﬁcient largely maintains performance relative to CheXfull and outperforms CheXrandom across lexical and clinical metrics (e.g., RadGraph, CheXbert-based Macro-F1). It remains competitive with state-of-the-art models trained on datasets 50× larger.

Significance and Claims

The paper claims that selective data curation is a practical and effective strategy for the resource-efficient development of medical foundation models. Key contributions include:

Challenging the "Scale" Paradigm: The work demonstrates that generalization ability does not necessarily scale predominantly with dataset size; instead, optimizing which data is used can be more influential than increasing how much data is used.
Addressing Imbalance: By prioritizing under-represented samples, the method mitigates the bias toward common findings inherent in standard pretraining, improving robustness for rare thoracic conditions.
Practical Accessibility: By reducing computational requirements by over 70%, the approach lowers the barrier to entry for academic and clinical institutions lacking industrial-scale resources, enabling the development of high-performing foundation models without massive infrastructure.
Generalizability: While focused on CXRs, the authors suggest the strategy is extensible to other medical imaging modalities and multimodal datasets.

The authors maintain a modest tone, acknowledging limitations such as the focus on a specific encoder architecture (DINOv2), the exclusion of certain rare diseases with extremely low sample counts (where curation alone is insufficient), and the need for prospective clinical validation before deployment. They emphasize that their findings support a shift toward data-centric strategies where quality, composition, and utilization are prioritized alongside scale.