TOPLINE:

According to a new study, the use of an artificial intelligence-based "You Only Look Once (YOLO)" algorithm showed high precision in detecting B lines on lung ultrasound (LUS) images, with substantial agreement being found with expert annotations.

METHODOLOGY:

• Researchers conducted an observational agreement study in Italy and analysed 386 images from 46 patients using anonymised internal and clinical online databases Butterfly and Grepmed.
• LUS frames were acquired using a 14-zone method comprising seven zones on each side of the lung, with expert sonographers being blinded to YOLO detection results and identified B lines through a rectangular region of interest.
• The implementation of the YOLO-based model included a fivefold cross-validation approach, with 80% of images being used for training and 20% being used for validation and testing to avoid overfitting.
• Performance evaluation metrics included weighted kappa statistics to assess the agreement between the algorithm and expert annotations, as well as precision (the proportion of correctly detected B lines among all detections), recall (the proportion of actual B lines accurately detected), and F1 scores (a weighted mean of precision and recall).

TAKEAWAY:

• The weighted kappa was 0.68 (95% CI, 0.64-0.72), indicating substantial agreement between the algorithm and expert annotations.
• The YOLO-based model achieved a precision rate of 0.92 (95% CI, 0.89-0.94), a recall rate of 0.81 (95% CI, 0.77-0.85), and an F1 score of 0.86 (95% CI, 0.83-0.88).
• In total, 352 true positives and 31 false positives were identified across validation sets, with 83 false negatives.

IN PRACTICE:

"The proposed YOLO-based algorithm has demonstrated its potential to significantly enhance diagnostic support by accurately detecting B-lines in LUS images," the authors wrote. "These results collectively suggest that the YOLO-based model offers a strong balance between precision and recall, achieving performance metrics comparable to or exceeding those reported in the literature," they added.

SOURCE:

This study was led by Alberto Bottino, University of Salento, Lecce, Italy. It was published online on September 11, 2025, in the Journal of Ultrasound.

LIMITATIONS:

The grid-based detection architecture presented challenges when objects were positioned close together, particularly with multiple clustered B lines. Low-quality images were excluded, which may have removed challenging cases from the samples.

DISCLOSURES:

Open access funding was provided by Consiglio Nazionale Delle Ricerche. Two authors reported being shareholders of Amolab S.r.l., a National Research Council spin-off company that may or may not have benefited from the results of this study.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.