Lung ultrasound interpretation using deep learning for the detection of B-lines in dogs.

A human-trained deep learning lung ultrasound model accurately detected B-lines in dogs, supporting AI cross-species diagnostic utility.

This study evaluated the performance of a deep learning (DL)-based B-line detection algorithm (BLDA), originally trained on human lung ultrasound (LUS) images, when applied to canine LUS data. A dataset of 1,950 ultrasound clips from 201 LUS examinations across 90 dogs was assembled from 4 separate studies. Clips were labeled as A-line or B-line profiles by an expert reviewer and used to assess the algorithm's diagnostic performance. A calibration step was performed to optimize clip-level classification before evaluation on a held-out test set. The BLDA demonstrated an overall accuracy of 87%, with a sensitivity of 73% and specificity of 93% for B-line detection. Zero-shot performance (without any canine-specific training) was strong, yielding an area under the curve (AUC) of 0.96. The algorithm performed best when distinguishing strong versus weak B-line profiles, with weaker profiles posing greater classification challenges. Gradient-weighted heatmap analysis confirmed that the model focused on image regions plausibly relevant to LUS interpretation, supporting mechanistic validity. These findings demonstrate meaningful cross-species generalizability of a human-trained DL model to dogs, without requiring species-specific retraining. This has significant implications for accelerating AI-assisted diagnostic imaging tools in veterinary medicine, particularly for point-of-care identification of pulmonary edema and other B-line-associated conditions in small animal patients.

This summary was distilled by AI and may occasionally misinterpret data. Confirm critical details with the primary literature before clinical application.