BIOANALYSISPAPER
Across 512 ovarian ultrasounds, a 2026 study found the top single AI model beat expert radiologists. The strongest score, though, came from combining them.
A peer-reviewed comparison of sixteen deep-learning models against expert radiologists reading 512 ovarian ultrasounds found the best single AI model beat the experts on the study's own data. The strongest workflow, though, was human plus machine.
The paper, [published this year in the Journal of Ultrasound in Medicine](https://pubmed.ncbi.nlm.nih.gov/42572015/), compared radiologist scoring under the American College of Radiology's 2022 ovarian risk system, called O-RADS v2022, against sixteen different deep-learning architectures trained on the same images. The dataset was 512 ultrasound images from 227 patients at a single center; 110 of those patients had at least one malignant lesion. The question was not whether AI could match a radiologist's eye. It was whether either could match both of them working together.
Radiologists using O-RADS v2022 alone, the standard clinical workflow today, scored 68.0% accuracy and 0.683 on AUC, a standard measure of diagnostic accuracy where 1.0 is perfect and 0.5 is a coin flip. The convolutional neural networks, or CNNs, ranged from 59.2% to 86.4% accuracy, with AUCs of 0.620 to 0.908. That spread matters: there is no single "AI score." A bad architecture underperforms a trained radiologist; a good one outscores one. Performance is model-dependent.
The top single model was ViT16-384, a type of image-recognition AI originally developed for general computer-vision tasks, retrained on this dataset. It hit 87.4% accuracy and an AUC of 0.941, the highest score in the study. A Vision Transformer (ViT) processes an image in patches rather than scanning it top-to-bottom like a CNN, which tends to help on subtle texture and shape cues. That mechanism helps explain why the lead changed architectures on this task.
But the paper's headline result is the hybrid. When the radiologist's O-RADS v2022 risk score was combined with the AI's prediction, accuracy rose significantly for 9 of 12 CNNs and 3 of 4 ViTs, the authors report. The framework being tested is not replacement. It is a calibration step: the AI flags a region, the radiologist weighs it against the patient's history and the rest of the scan, and the two scores are written together. On ultrasound, where reader-to-reader variability is a known clinical problem, that combination smooths the score.
The study is single-center and retrospective: 227 patients at one institution, with images labeled from that institution's clinical record. There is no prospective replication, no multi-site cohort, and no claim of FDA review or clinical guideline change. The authors' own conclusion, in the paper's abstract, is that hybrid human-AI integration yields the highest accuracy for standardizing ultrasound interpretation and reducing false-positive diagnoses, a careful statement that AI helps, not that AI decides.
What the result actually does, and what most news coverage of medical AI tends to skip, is map the gap between "AI beats radiologists" and "AI plus radiologist beats radiologists." A 0.683 AUC on a system already in clinical use is also a number worth noticing. It means a meaningful slice of adnexal lesions, masses in the ovary or the nearby fallopian-tube region, get a borderline score under the current standard. The hybrid finding is what makes that borderline score legible.
The team's next step is the part worth watching. The paper's open questions, in the discussion section, are multi-site validation and a prospective trial. Until those land, the working read is that the radiologist's eye still leads, and the AI's job is to push that eye toward the same answer twice in a row.