The Study On Deep Learning Algorithm And Application In Medical Image Analysis

Objective:In the development of smart medical care,medical imaging analysis systems that are in urgent need of intelligence are needed to assist radiologist to improve diagnostic efficiency and accuracy.Although the existing medical image analysis research work has made great achievements and progress,it still does not meet the increasing clinical needs of the existing clinics or cannot make the clinicians truly satisfied,and still has a certain distance for the clinically practical stage with accurate and reliable.Intelligent medical image analysis still has challenges and difficulties stemming from data and algorithms.In order to overcome the difficulty of medical image analysis and realize accurate computer-aided diagnosis,this paper is devoted to medical image analysis based on deep learning and proposes practical and generalized algorithm models for the difficult problems to be solved.Methods and results:A structured deep learning model was proposed for the multi-classification problem of breast cancer pathological images.A full-scale convolutional detection of deep multi-scale multi-task was proposed for MRI-based detection of spinal lesions.Aiming at the problem of segmentation and diagnosis report generation of multi-structured lesions of the spine,a new Recurrent Generative Adversarial Network was proposed.(1)A structured deep learning model CSDCNN is proposed for fine-grained image classification.The innovation is that it can not only automatically adjust the feature space distance between different categories,but also extract the high-level semantic features of the image to overcome the shortcomings of traditional machine learning methods in feature extraction.By verifying the multi-classification problem of breast cancer pathological images by CSDCNN,the first high-precision and fully automatic breast cancer pathological image multi-classification method was realized,which can subdivide the pathological image of breast cancer into eight categories.CSDCNN achieves an average of 94%multi-class accuracy in the validation set and testing set.(2)A deep multi-scale multi-task full convolutional network(DMML-Net)is proposed for the detection and classification of lesion targets.The innovation of this method is that DMML-Net combines multi-scale,multi-output learning and multi-task regression learning theory,with strong multi-scale multi-task learning ability,as well as generalization and robustness.By applying DMML-Net to the disease detection and classification tasks in spinal MRI,the first diagnostic mechanism for lumbar intervertebral foramen stenosis based on pathogenesis was realized.DMML-Net achieves 90%detection accuracy and has good potential for clinical application.(3)A new Recurrent Generative Adversarial Network(Spine-GAN)was proposed for the diagnosis of spinal diseases.The purpose is to use MRI to automatically segment and classify various spine structures such as intervertebral discs,vertebrae,and neural foramen,and then automatically generate a complete clinical diagnosis report.The innovation of this method is that Spine-Gan can not only overcome the high variability and high complexity of the spine structure in the MRI image but also preserve the subtle differences between the normal spine structure and the abnormal structure dynamically.The spatial and pathological relationship between the adjacent structures is not obvious but important,thus breaking through the limitations of small data sets.Spine－GAN achieved accurate segmentation of three spinal diseases,radiological classification,and pathological correlation.Specifically,Spine-GAN gets a pixel accuracy of 96.2%,a specificity of 89.1%and a sensitivity of 86%.Conclusion:In this paper,we propose three innovative deep learning algorithms for the automatic diagnosis of breast cancer pathology images,the automatic localization,and classification of spinal disease lesions,the segmentation of spinal lesions and the generation of radiology diagnostic reports.All three algorithms have reached a high level of diagnosis and have potential and value for clinical application.