An Algorithm for Recognition of Left Atrial Appendage Boundaries in Echocardiographic Images

Introduction: Atrial fibrillation is the most common cardiac arrhythmia of clinical significance and is associated with increased risk of stroke through thrombus formation primarily in the left atrial appendage (LAA). LAA occlusion using a purpose-built device is a growing procedure. In this study, we aimed to develop a computer aided diagnostic system for recognition of the LAA in echocardiographic images.Material and methods: 3D echocardiographic images of the LAA of 26 patients who were successfully treated with an LAA occluder were used in this study. 208 3D derived 2D images in the axial plane were derived from each 3D dataset. 562 images in which the LAA boundaries were highly recognizable were selected. The proposed convolutional neural network (CNN) in this study is based on an open source object identification and classification platform that is compiled under the YOLO algorithm. 419 and 143 images were used for training and testing the algorithm respectively. Results: Algorithm performance on identifying the LAA region on the set of 143 images was compared to the traced regions on the same images by the expert in echocardiography using intersection over union (IOU) algorithm. The algorithm was able to correctly identify the LAA region in all 143 examined images with average IOU of 90.7%. Conclusion: The proposed image based CNN algorithm in this study showed high accuracy in recognizing LAA boundaries in the echocardiographic images. The method can be used in developing algorithms for automated analysis of the area of the LAA that is used for device sizing and procedural planning in the LAA occlusion procedures.