Research On Auxiliary Diagnosis And Treatment Of Congenital Heart Disease Based On Deep Learning

With the popularization of medical equipment in the medical field and the development of computer-aided diagnosis technology,the auxiliary diagnosis and treatment of diseases has become a popular research field in recent years.At present,the detection of lesions for various diseases has achieved good diagnostic results.However,there are few diagnostic algorithms for congenital heart disease(CHD)and there does exist some technical challenges,such as multi-modality data of CHD have multi-view images,and there also have a great difficulty in obtaining markers of suspected lesions in CHD,leading to multi-modality data processing and multi-view learning in the diagnosis of CHD.To tackle these questions,we combine deep learning technology with computer-aided diagnosis technology to diagnosis CHD disease.Our research contents include: location of suspected lesions,multi-view feature learning and multi-modality data fusion.The main contributions are as follows:(1)To well detect the location of lesions,we apply single shot multi-box detector(SSD)to the diagnosis of CHD.A refined-SSD based lesion detection network is proposed from the perspective of the characteristics of CHD lesions.The redesign of the default frame and the input of continuous video frame images strategies optimize the detection performance of the detection network for the lesion area,and further effectively solve the location of lesions.(2)For the classification of suspected lesions in CHD,we propose a multi-view fusion based congenital lesion detection algorithm,which utilizes a multi-view fusion algorithm to fuse multi-view information to obtain a more comprehensive feature representation for the identification of abnormal types of lesions.Finally,the identification of lesion areas in all directions is achieved.(3)For the multi-modality data processing problem,we propose a motionappearance compensation algorithm for the fusion of different modal data.The diagnosis of CHD contains more auxiliary diagnostic information like the blood flow of color Doppler in which the blood flow information is superimposed on 2D ultrasound structure images.Under such circumstance,directly applying existing image analysis methods could suffer overfitting easily due to the large variance of information from different data sources.The motion-appearance compensation algorithm can well fuse the 2D structural features with the high-speed blood flow,so the acquired features could pay more attention to the area of the suspected lesion area in the features,and provide a refined feature representation for diagnosis of CHD.The experiment on the diagnosis of CHD show that the proposed algorithms achieve more accurate detection results when compared with the traditional lesion detection algorithm,and also illustrate the effectiveness of our proposed algorithms.