Artificial Intelligence Diagnosis Analysis Of Cardiac Patent Foramen Ovale

Patent Foramen Ovale（PFO）is one of the most congenital heart disease in the world.Recent studies show that PFO has close relationship with ischemic/cryptogenic stroke,migraine and decompression sickness caused by paradoxical embolism.Timely screening and effective diagnosis are of great significance for the prevention,diagnosis and treatment of this disease group.With the development of artificial intelligence,machine learning has been widely used in various medical image analysis tasks and achieved excellent results and showed great potential.However,due to the complexity and semi-invasive nature of conventional diagnostic methods,image noise and motion artifact interference,the complexity of bubble detection and limited quantity and poor quality of datasets,PFO has faced great challenges.Most of the existing methods exist in clinical practice,and there is little substantial progress in engineering exploration and research.Therefore,we explore the possibility of diagnosing PFO disease based on artificial intelligence methods,and non-invasive and reliable intelligent diagnosis methods.In this study,a total of 200 dataset（144 positive and 56 negative cases）were collected by TTE（transthoracic echocardiography）and cTTE（contrast transthoracic echocardiography）,and double-blind diagnoses were performed by five physicians with varying experience.In order to deal with the difficulties and challenges of PFO,we transformed the diagnosis of PFO into bubble detection task based on clinical diagnosis experience and key diagnostic basis other than clinical gold standard,and proposed two-stage radiomics analysis based on spatio-temporal information and object detection analysis based on deep learning methods.Firstly,aiming at the interference and bubble detection difficulties in PFO diagnosis,in the time domain,a method combining multi-mode image and prior knowledge of cardiac cycle was proposed to remove the noise and interference in echocardiography;in the spatial domain,a superpixel clustering segmentation method was proposed for bubble coarse-grained classification and a radiomics method was utilized for bubble secondary fine-grained classification.These methods make full use of temporal and spatial dimension information to perform bubble detection and achieve automatic PFO diagnosis.In 200 cases dataset of 144 positive and 56 negative cases,the classification results were 0.7750 accuracy,0.7847 sensitivity,0.7500 specificity and AUC 0.8063.Further,based on the experience and limitations of the first work,combined with a small amount of image labeling and pre-training model,Faster R-CNN object detection method was proposed to classify diseases by left atrium detection,patient-by-patient image frame-by-frame bubble detection,and voting classification mechanism.In 67 cases dataset of 37 positive and 30 negative cases,the bubble detection results were 0.8950 recall and 0.7795 average precision and the classification results were 0.8657 accuracy,0.8378 sensitivity and 0.9000 specificity.The above two works respectively conducted machine learning and deep learning method to perform preliminary attempts to PFO diagnosis and obtain potential diagnosis results.In the future,it is expected to improve the diagnostic accuracy and efficiency of PFO,assist doctors for diagnosis,and promote early screening of such diseases.