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

Abstract Background Deep learning (DL) shows promise for interpretation of lung ultrasound (LUS) images in humans, but its performance in animals remains underexplored. Hypothesis/Objectives Assess performance of a B-line detection algorithm (BLDA) trained on LUS images from humans when applied to images from dogs. Animals A total of 1,950 clips collected from 201 LUS examinations in 90 dogs across 4 studies. Methods Ultrasound clips were collated and labeled with A-line or B-line profiles by an expert reviewer. A DL model previously trained on LUS images from humans was applied to detect presence of B-lines at a framewise level. A clip classification algorithm was calibrated to maximize clip-level performance of the BLDA using a calibration data set. Performance of the DL model and BLDA was assessed on a held-out test set of LUS images. A heatmap-based explainability method was used to visualize regions most utilized by the model for predictions. Results When applied to the test set, the BLDA showed an overall accuracy of 87%, with sensitivity of 73% and specificity of 93%. The algorithm performed best when identifying images with strong (versus weak) B-line profiles. Assessing raw model predictions, zero-shot performance for identification of B-lines was excellent (area under the curve, 0.96). Heatmaps suggested that the DL model utilized image areas that were plausibly relevant for LUS interpretation. Conclusions and clinical importance A LUS model trained on images from humans maintained strong performance when applied to dogs, supporting cross-species generalization to accelerate veterinary diagnostic innovations.