CXRBase

Two chest radiographs from two very different patients look almost identical at a glance: the same ribs, the same mediastinal silhouette, the same diaphragmatic contours. What separates them is often a few centimetres of opacity in one lung field. That macroscopic sameness is precisely why masked image modelling suits radiography — filling in a hidden patch of a chest film requires an internal model of what normal thoracic anatomy is supposed to look like there — and also why it is delicate, since the pathological signal is small and local enough that a random mask can simply delete it.

CXRBase takes that bet directly. Developed at the Centre for Perceptual and Interactive Intelligence at the Chinese University of Hong Kong with Shanghai AI Laboratory and SenseTime Research, and posted as a preprint in October 2024, it is a masked autoencoder trained to reconstruct chest X-rays from versions with 75% of their patches removed. Pretraining is applied consecutively — first on natural images from ImageNet-1k, then on a pooled corpus of 1.04 million unlabelled radiographs — and runs as a single 800-epoch job whose best epoch is kept as one checkpoint. Everything downstream reuses that frozen starting point.

It is deliberately an image-only model, which places it opposite the report-supervised line of chest radiograph work: CheXzero, CXR-CLIP and Google's CXR Foundation all learn from radiographs paired with their written reports and can be queried with text. CXRBase never sees a report and offers no zero-shot text-prompted classification; adaptation means fine-tuning the encoder with task labels. The trade is that it needs no paired text at all, so any archive of raw images becomes usable pretraining data.

#Key Features

• Single checkpoint, many cohorts: One pretraining run produces the encoder that is then adapted to roughly twenty downstream datasets, including five multi-centre private cohorts absent from the pretraining corpus.
• Label-efficient adaptation: Classification is evaluated in a few-shot setting and localization at 50 labelled examples per disease, the regime where pretraining is supposed to pay off.
• Screening and localization from one backbone: The same features support binary negative screening, 14-label disease classification, and bounding-box localization of 15 chest findings.
• Encoder-only downstream use: The reconstruction decoder is discarded after pretraining; only the ViT-large encoder plus a multilayer perceptron head carries into clinical tasks.

#Technical Details

The encoder is a ViT-large with 24 transformer blocks and 1,024-dimensional embeddings; the reconstruction decoder is a much lighter ViT-small with eight blocks at 512 dimensions. Images are cropped to the chest region, resized to 512×512, then randomly cropped to 224×224 and tokenized into 16×16 patches, with a 0.75 mask ratio. Pretraining used a batch size of 2,048 across eight GPUs for 800 epochs, the first 30 spent warming up the learning rate. The 1.04-million-image corpus is 69.5% public — MIMIC-CXR, PadChest, CheXpert, ChestXray14 and CXR-AL14 — and 30.5% in-house.

On negative screening, CXRBase averages 79.1% AUC across five public test sets and 82.7% across five private ones. For single-disease screening it reaches AUROC of 99.5% on SIRM, 96.7% on StonyBrook and 99.2% on RICORD for COVID-19, 81.6% on RSNA Pneumonia, and 99.9% and 99.8% on the TBX11K and TB Chest X-ray tuberculosis sets. Few-shot multi-label classification averages 76.9% AUC and 35.6% F1 on public data and 84.3% AUC and 37.9% F1 on the five private cohorts. Localization is the harder task and the numbers say so: 19.1% mean AP50 on public datasets and 19.2% on private ones under the 50-shot setting.

#Applications

The intended use is a starting point for hospitals that hold large unread radiograph archives but little annotation budget. Fine-tuning the shared encoder gives triage-oriented normal-versus-abnormal screening, multi-label reporting support across common thoracic findings, and coarse lesion localization for reader assistance — each from a modest number of labelled studies rather than a purpose-built dataset. Evaluation on prospectively and retrospectively collected cohorts from five separate Chinese hospitals is aimed at the generalization gap that limits site-specific radiography models.

#Impact

CXRBase is a clean demonstration that masked autoencoding scales to radiography at the million-image mark and transfers to cohorts and diseases outside its pretraining distribution, including COVID-19 and tuberculosis datasets that postdate or sit entirely apart from the public corpora it learned from. Its practical reach, however, is limited by what has actually shipped. The work remains a single-version arXiv preprint with no peer-reviewed publication, its five external validation cohorts are private, and the GitHub repository its Code and Data Availability statement names as the home of the code, the pretrained models and a newly released dataset does not resolve — no weights, source or data have been made available. Readers can learn from the recipe and the evaluation design, but cannot reproduce or deploy the model as released.