Viewing Breast Ultrasound Images in Sequence: AI Assistance Improves Junior Physicians’ Diagnostic Accuracy

On breast ultrasound, benign masses and malignant lesions sometimes show similar contours and echoes, making interpretation a test of experience. A prospective multicenter study published on October 10 in *Holistic Integrative Oncology* showed that having AI analyze dynamic scanning videos and then provide its findings for physicians’ reference could improve less experienced physicians’ ability to distinguish benign from malignant lesions.

The study focused on lesions classified as categories 4 to 5 under the Breast Imaging Reporting and Data System (BI-RADS), meaning cases already assessed on imaging as suspicious or highly suggestive of malignancy. The research team collected data from four hospitals between November 2019 and January 2023, including 532 lesions overall, of which 334 were used for model training and validation. All benign or malignant diagnoses were confirmed by pathology following biopsy or surgery.

The model, called TGMIL, analyzes images from different time points in a scanning video in sequence and assigns greater weight to frames containing more diagnostic information. Compared with selecting just one representative image, this design aims to retain the morphological changes seen as the probe moves across a lesion, helping address features that are difficult to characterize in a single frame.

The published paper tested the model using 120 patients and 198 videos, reporting a patient-level area under the curve (AUC) of 0.845. This metric measures the ability to distinguish benign from malignant lesions; it does not mean an accuracy of 84.5%. With AI assistance, diagnostic accuracy increased from 66.7% to 89.2% in the junior physician group and from 80.8% to 90.0% in the attending physician group. A summary of the same study published by 生物通 on October 12 also presented these patient-level results.

The unit of comparison for these figures requires attention. An early preprint made public on Research Square on May 21 reported, using videos as the unit of analysis, that junior physicians’ accuracy increased from 69.2% to 85.9%. The published version instead evaluated results after aggregating multiple videos for each patient. The two sets of results therefore cannot be used interchangeably, nor can the difference be interpreted as another improvement in the model over a few months.

The evidence still comes from a specific population undergoing diagnostic assessment at hospitals. The study was limited to BI-RADS category 4 to 5 lesions larger than 5 millimeters and excluded cases with poor image quality, among other conditions. It did not cover the general asymptomatic screening population. The paper also noted that overlapping imaging features of benign and malignant lesions still lead to misclassification. When AI recommendations conflict with senior physicians’ judgments, they may also interfere with decisions that were initially correct.

These results support further validation of AI’s value as an interpretation aid, but they have not yet demonstrated that biopsies can safely be reduced on this basis. Although the early preprint’s title mentioned the possibility of reducing biopsies, improved interpretation accuracy does not establish that fewer tests can be performed without increasing missed diagnoses. The next step requires evaluating how the tool affects testing decisions and patient outcomes in a broader range of patients and actual clinical workflows. Pathological confirmation and comprehensive clinical assessment remain essential.