The Research And Application Of Deep Learning Algorithms For Medical Image Segmentation And Estimation

Objective:Using image segmentation and measurement technology to obtain valuable information from medical images is a research hotspot in the field of artificial intelligence.In recent years,the rapid development of medical imaging technology has brought technical challenges to intelligent computer-assisted diagnosis.Artificial intelligence technology based on machine learning algorithms,especially the deep learning algorithms,has been gradually applied to the medical image segmentation and measurement.However,there are still many problems have not been completely solved for computer-aided diagnosis,such as the segmentation of complex tissue structure areas,the indices measurement,and the disease-assisted screening.Based on the deep learning algorithm,this paper has made a study on the measurement of left ventricular ejection fraction based on cardiac MRI images,the segmentation of multiple spinal structures from 3D MRI images,and the assisted screening of Corona Virus Disease 2019(COVID-19)from X-ray images.Finally,this paper proposed three reliable deep learning algorithms with excellent performance.Methods and results:(1)This paper proposes a cardiac cycle feature learning(CCFL)for the measuring of left ventricular ejection fraction(LVEF)from cardiac MRI images.CCFL algorithm constructs of a cardiac cycle extraction module(CCE),a motion feature fusion and extraction module(MFFE),and a fully connected regression module(FCR).The CCE module can generate optical flow fields and analyze them to obtain a cardiac cycle from a series of cine MRI images;the MFFE module achieves feature extraction and builds a time-series modeling from the end-systole and end-diastole images;and the FCR module can regress the LVEF from the extracted feature maps.Experiments prove that our proposed CCFL remarkably outperforms state-of-the-art algorithms.It achieves a correlation coefficient of 0.946,a mean absolute error of 2.67,and a standard deviation of 3.23.(2)This paper proposes an adversarial network called S~3eg ANet for segmenting multiple spinal structures from MRI images.A multi-modal autoencoder module(MMAE)was leveraged to extract fine-grained structural information from complex 3D spinal structures.Besides,S~3eg ANet uses a cross-modality voxel fusion module(CMVF)fusing comprehensive spatial information of multi-modality MRI images.A multi-stage adversarial learning strategy(MSAL)is proposed for achieving high accuracy and reliable segmentation of multiple spinal structures simultaneously.The S~3eg ANet achieves average Dice coefficient of 88.3%and average sensitivity of 91.45%.(3)This paper proposes a discriminative cost-sensitive learning(DCSL)algorithm for screening COVID-19 based on X-ray images.DCSL combines the advantages of fine-grained classification and cost-sensitive learning.First,DCSL designed a conditional central loss(CCL),which learned a deep discriminant representation.Secondly,DCSL has established score-level cost-sensitive learning,which can adaptively adjust the substitute value of COVID-19 cases misclassified into other categories.This paper collected a multi-class data set with a total of 2239 X-ray images,including 239 confirmed cases of COVID-19,1000 confirmed cases of bacterial or viral pneumonia,and 1000 healthy people.Extensive experiments show that the DCSL algorithm is significantly better than the state-of-the-art algorithm.Its achieves a accuracy of 97.01%,a sensitivity of 97.09%,and an F1-score of 96.98%.Conclusion:In this paper,three deep learning algorithm models are designed,which focus on the measurement of LVEF,segmentation of 3D spinal structures,and the auxiliary screening of COVID-19.Among them,CCFL can be used as a measuring tool of LVEF;the S3eg ANet can finely segment the spinal structures from MRI images;the DCSL algorithm can achieve auxiliary screening of COVID-19 based on X-ray images.