Ultrasound labels don't need every frame: EvoDiff generates diffusion probes, then sends only retrieved real images to experts
Synopsis
EvoDiff recasts active-learning sample selection as generate-then-retrieve: diffusion sampling in feature space produces synthetic probes, and only the real unlabeled ultrasound images those probes retrieve are sent to experts for annotation. Across 15 datasets and 16,696 images spanning six organ tasks, it beat random sampling and several active-learning baselines in most settings on both ResNet-50 and the USFM foundation model, and a model-based annotation-effort metric estimated a 24% to 84% workload reduction relative to random sampling.
Interpretation
Training labels for ultrasound diagnostic models no longer come from uniform sampling; generative probes retrieve higher-information real images from the unlabeled pool before annotation. Earlier active learning was mostly designed for a single task and a single model family, ranking samples by predictive uncertainty or feature representativeness, and on heterogeneous, heavily imbalanced ultrasound cohorts such strategies tend to favor the majority class or redundant samples. Across 48 dataset-and-budget settings, eight internal datasets, six organ tasks and six cumulative annotation budgets from 5% to 50%, EvoDiff ranked first in accuracy in 43 and second in the remaining five, with a largest gap of 0.88 percentage points from the best method.
Selection does not rely on the downstream classifier's posterior; it first generates synthetic samples worth inspecting in feature space, then uses them to match real images. Diffusion models had mostly been used for medical image synthesis rather than to guide active-learning queries. Synthetic images serve only as retrieval probes and are neither annotated nor used in downstream training; one acquisition round on BUS-BRA at the 10% cumulative annotation budget took 6.8 GPU-hours on a single NVIDIA A40, versus 19.4 seconds for Max-Entropy and 86.5 seconds for BALD.
Perspective
The evaluated scope is six organ diagnostic tasks on 2D B-mode ultrasound images, with eight internal and seven external cohorts and ResNet-50 and USFM as downstream models; the 24% to 84% annotation-workload reduction is a model-based estimate from the HAE metric, which multiplies annotation counts at matched performance by per-image annotation times measured in the reader study.
The authors note that the framework relies on a frozen feature encoder during generation and selection, so embedding shifts may affect sample ranking; validation is limited to 2D B-mode ultrasound, with extension to MRI and CT left for future work; acquisition-side compute cost is markedly higher than uncertainty-based methods; and the 24% to 84% reduction assumes the same unit annotation time for both strategies. In addition, BUSI lacks patient identifiers in its public release, so only image-level partitioning was possible and strict patient-level separation cannot be verified.