Using generated ultrasound images to choose real ones worth labeling

Ultrasound archives can contain many images without the diagnostic labels needed to train artificial intelligence. An examination may produce numerous overlapping frames, while its report describes the case as a whole. Connecting individual images to diagnoses takes expert work. Could choosing those images more carefully reduce the labeling burden?

An arXiv preprint by Yaofei Duan and colleagues introduces EvoDiff, a system that uses generated ultrasound images to help decide which real images deserve annotation. Its central distinction matters: the artificial images guide a search. They are neither labeled by radiologists nor used to train the downstream diagnostic models.

Generate, then retrieve

EvoDiff combines active learning—selecting examples for annotation rather than accepting a random sample—with a diffusion model. Such models learn to reverse a process that adds noise to images, allowing them to generate new examples.

The researchers first train the generator and a guidance classifier using images identified by organ, then fine-tune them using a smaller set with diagnostic labels. During generation, these networks remain fixed. The system changes the image being generated, not the networks’ parameters.

Its guidance uses predictive entropy, a measure of how uncertain the classifier is about a diagnosis. An evolutionary search repeatedly selects, combines and perturbs candidate image states, favoring uncertainty while penalizing departures from the learned category features. The practical uncertainty score is a substitute for an ideal Bayesian information-gain calculation, not a direct measurement or guarantee of how much a diagnostic model will learn.

The resulting synthetic images become retrieval probes. EvoDiff compares their numerical image representations with those of real, unlabeled scans and retrieves similar examples for expert annotation. Across successive rounds, stronger guidance is intended to move selection from clearer cases toward more ambiguous ones—a training curriculum rather than an immediate concentration on the hardest images.

What the benchmarks showed

The study included 16,696 images across 15 datasets, covering appendix, breast, carotid artery, liver, ovary and thyroid tasks. Eight datasets supported training and internal testing; seven were external evaluation cohorts. The researchers tested both conventional ResNet-50 classifiers and USFM, a pretrained ultrasound foundation model. The ResNet-50 experiments used a separate diagnostic model for each dataset, not one universal multi-organ classifier.

Each internal dataset had a fixed split: 80% formed the candidate pool, 10% supplied initial diagnostic labels and 10% was reserved for testing. Reported acquisition budgets describe fractions of that candidate pool annotated in addition to the initialization set. Thus, a 10% budget does not mean training used only 10% of all available images.

For ResNet-50, EvoDiff ranked first in accuracy in 43 of 48 dataset-and-budget comparisons against other acquisition methods, and second in the remainder. At the 50% candidate-pool budget, fatty-liver accuracy reached 97.24%, compared with 97.41% for fully supervised training. Appendix accuracy likewise approached its fully supervised reference, but both remained around 64%—a reminder that approaching a reference does not establish clinical usefulness.

Foundation-model results were less uniform than the paper’s broad narrative suggests. Supplementary results show, for example, that USFM’s balanced accuracy on the internal thyroid dataset at the 10% budget was 60.08% with EvoDiff versus 67.37% with random selection. Balanced accuracy gives equal weight to performance on each class, making it useful when one class dominates. Such exceptions mean EvoDiff cannot be described as consistently improving every model and task.

External results were more encouraging: EvoDiff-selected source data improved USFM’s balanced accuracy in 42 of 45 source-to-target-and-budget comparisons, without additional training on the target cohorts. However, the largest reported relative gains were selected after comparing budgets, rather than representing average improvements.

Fewer labels, not proven time savings

The headline workload reduction—24% to 84%—is a model-based estimate. Six radiologists supplied annotation timing and accuracy measurements. The researchers combined these with the annotation counts estimated to reach matched downstream performance, interpolating between random-sampling results where necessary.

Crucially, the calculation assumes the same average time per correctly annotated image for both selection strategies. It therefore does not directly demonstrate that radiologists completed an EvoDiff-selected workload that much faster. Selected difficult cases could take different amounts of time.

Computing costs also increase sharply. One representative acquisition round required 6.8 GPU-hours on an NVIDIA A40, compared with 19.4 seconds for maximum-entropy selection. That work runs offline, but it remains a substantial resource trade-off.

Other limits include validation only on two-dimensional B-mode images, unverifiable patient separation in one public breast dataset, and five training folds evaluated on the same fixed test set rather than independently repeated dataset splits. Sensitivity gains were not uniform and sometimes accompanied lower specificity. The preprint supports further investigation of selective annotation, not a conclusion that the resulting systems are clinically safe.