BUSGen

Ultrasound is the workhorse of breast imaging — non-invasive, inexpensive, and more sensitive than mammography in dense or younger breasts — but it has resisted the data-hungry deep learning approaches that transformed chest radiography and histopathology. Public breast ultrasound datasets are tiny, because the images are patient-identifiable and legally encumbered, so models trained on them overfit, generalise poorly across scanner vendors, and rarely see enough rare pathology to learn it.

BUSGen attacks that scarcity from the supply side. Rather than a discriminative encoder that must be paired with labelled data, it is a conditional diffusion model pretrained on more than 3.5 million breast ultrasound images, which can then be adapted with a few real examples to manufacture unlimited, label-balanced, synthetic training sets for whatever downstream task is at hand. Classifiers and detectors trained purely on those synthetic images — the authors call them BUS-DMs — are what actually get deployed.

The model comes from a consortium led by Peking University with the Peking University Cancer Hospital & Institute, Peking Union Medical College Hospital, the Cancer Hospital of the Chinese Academy of Medical Sciences, Stanford University, and the imaging company Yizhun Medical AI. It appeared as a preprint in January 2025 and was published in Nature Biomedical Engineering in 2026.

#Key Features

• Conditional generation on clinical structure: Every sample is conditioned on a lesion bounding box encoded as positional embeddings, a benign/malignant pathology label, and the scanner device type as one-hot vectors, so generated images arrive with usable labels attached.
• Few-shot adaptation without touching the backbone: The pretrained weights stay frozen and only low-rank adapters are fine-tuned per task, which is what makes adaptation viable from small downstream cohorts.
• Pixel-space diffusion: The model denoises directly in pixel space rather than in a compressed latent, a deliberate choice to retain the high-frequency speckle and margin detail that carries diagnostic signal in ultrasound.
• Device-invariant synthesis: A CycleGAN-based augmentation step transfers each real image into all 19 scanner styles seen in pretraining, which suppresses the spurious device correlations that make ultrasound classifiers brittle across sites.
• Privacy-preserving output: Sampling is constrained to avoid reproducing training images, so the generated repositories can be shared as fully de-identified data.

#Technical Details

BUSGen is a conditional denoising diffusion probabilistic model with a deliberately lightweight U-Net of roughly 50 million parameters, trained over 500 diffusion steps at 160×160 resolution with classifier-free guidance. Pretraining used BUS-3.5M, an in-house multi-institution corpus of 3,518,495 images — 1,130,843 without lesions and 2,387,652 with annotated lesion boxes — drawn from 5,907 examinations of 4,636 patients across 3,749 lesions and more than 30 pathological subtypes, spanning ages 11 to 84 and scanners from a dozen vendors. Expert annotations for pathology, lesion location, and device were extracted from clinical reports and verified by clinicians. Inference uses DPM-Solver++ to cut sampling to 50 steps, producing an image in about 2.14 seconds on a single RTX 4090.

Downstream models trained on generated data were benchmarked against CLIP-based foundation models transferred from 1.65 million biomedical image–caption pairs and then fine-tuned on real breast ultrasound. On early diagnosis the synthetic-data model reached an AUC of 0.900 against 0.846 for the CLIP baseline, and on axillary lymph node metastasis prediction 0.895 against 0.807. In a reader study against nine board-certified radiologists using BI-RADS 4A as the operating threshold, it improved average sensitivity by 16.5% at matched specificity (P < 0.0001). Scaling the generated corpus to one million images made synthetic data as effective as collected real-world data on comparable performance curves.

#Applications

The intended user is a clinical AI group with a well-defined breast ultrasound task and too few annotated cases to train on: lesion detection, triage of opportunistically detected lesions, benign–malignant classification, molecular subtype prediction, or nodal metastasis prediction. BUSGen converts that small cohort into a large balanced training set, and because the synthetic images carry no patient identity, the resulting corpora can cross institutional boundaries that real scans cannot. The same property makes it useful for building shared benchmarks and for stress-testing detectors against scanner variation.

#Impact

BUSGen is one of the more concrete demonstrations that generative pretraining can substitute for data collection in a regulated clinical imaging domain, and the reader study is the strongest form of that claim: models trained on nothing but generated images outperforming experienced clinicians on early diagnosis. The evidence remains retrospective and largely single-consortium, with the pretraining corpus drawn from Chinese hospitals and unable to be released for privacy reasons, so external generalisation beyond the reported held-out sites is not yet established. Training and sampling code is on GitHub under an MIT licence, together with a toy configuration that reproduces the pipeline on the public BUSI dataset; the full pretrained model is reachable through the authors' online demo.