The invention provides a deep learning atrial septal defect detection method and device, and the method comprises the steps: obtaining an ultrasonic cardiogram, preprocessing the ultrasonic cardiogram, and extracting a region of interest; carrying out feature extraction on the region of interest, and identifying a subsword process atrial septal frontal section, a subsword process atrial septal sagittal section, a cardiac apex four-cavity cardiac tangent plane, a low-position paracarpal four-cavity cardiac tangent plane and a paracarpal aorta short-axis tangent plane; detecting the minimum distance point of the atrial septal defect; segmenting heart structures of a subsword atrial septal frontal section, a subsword atrial septal sagittal section, a cardiac apex four-cavity cardiac tangent plane, a low paracardial four-cavity cardiac tangent plane and a paracardial aorta minor axis section to obtain segmentation results, the heart structures including a left atrium, a right atrium, a left ventricle, a right ventricle, an aorta and pulmonary arteries; and according to the segmentation results, filtering the detected minimum distance point bounding box of the atrial septal defect, andtaking the filtered result as an atrial septal defect detection result.