Introduction: The detection of left atrial appendage (LAA) thrombi is one of the most frequent and challenging indications for transesophageal echocardiography (TEE). While automated image analysis by artificial intelligence algorithms has been evaluated for multiple purposes in transthoracic echocardiography, it has never been evaluated for TEE-based LAA thrombus detection.

Methods and Results: TEE studies of 77 patients (58 male, age 66 ± 13 years) with proven LAA thrombus were extracted from our database and anonymized. LAA thrombi were deemed to be present if at least two experienced TEE operators interpreted the studies to contain a thrombus and additional criteria such as low flow, positive contrast study, or sludge were present. In addition, TEE studies from 25 control patients (13 male, age 62 ± 9 years) were gathered. In total, 17.745 individual frames of the LAA in a 60° orientation and 11.792 individual frames of the LAA in a 120° orientation were analyzed by a convolutional neural network trained to classify images for the presence of LAA thrombus. In the 60° view, the algorithm had an accuracy of 85% in detecting LAA thrombus. In the 120° view, the algorithm's accuracy was 59%, most likely due to the increased morphological heterogeneity in this TEE view.

Conclusion: Automated image analysis by a convolutional neural network has the potential to support cardiologists in the detection of LAA thrombi. This may help to improve diagnostic accuracy in difficult or equivocal cases. Morphological variants of the LAA present challenges in a 120° view and warrant further studies to increase diagnostic accuracy.