Deep Learning Algorithm And Its Application In Medical Image Analysis

Objective: Deep learning is rapidly becoming the state of the art,leading to enhanced performance in various medical applications.The intelligent algorithm for medical image processing based on deep learning can not only automatically complete the extraction and selection of medical image features,but also construct new feature representations based on the latent texture structure of the image,its "end-to-end" representation and learning methods can be used in diagnosis play an important role.This paper is devoted to the research of high-performance deep learning algorithms for medical image processing and analysis,for the fine-grained multi-grading problem of intervertebral foramen stenosis,the diagnosis of multi-structural diseases in the global spine mode,and the problem of weakly supervised segmentation of skin cancer lesions,proposes three effective deep learning algorithm models for each task,respectively.Methods and results: Firstly,for the problem of automatic multi-grading MR images of foraminal stenosis,a simple and highperformance structured deep learning model is designed.Secondly,for the multiple tasks of spine multi-structure detection,contour segmentation and lesion classification/grading,the idea of joint learning between multi-structure and multi-task is proposed,and a progressive method for the diagnosis of spine multi-structure diseases under global mode Multi-task joint learning model.Finally,for the weakly supervised problem of skin cancer lesion detection and accurate segmentation,a weakly supervised skin cancer image segmentation algorithm based on superpixel region response is proposed.(1)Preoperative qualitative and graded diagnosis of intervertebral foramina stenosis is crucial for the clinician’s treatment strategy and patient health recovery.At present,doctors mainly adopt manual measurement and evaluation to diagnose intervertebral foramen stenosis,which is susceptible to doctors’ subjective factors and many other problems.To improve the accuracy of computer-aided diagnosis of intervertebral foramina stenosis and the diagnosis efficiency of doctors,this study proposes an automatic grading algorithm for intervertebral foramina stenosis based on deep learning.Firstly,we extracted the spinal foramen images from the sagittal spine MR image,and then these images are preprocessed.Secondly,a supervised deep convolutional neural network model is designed to achieve the automatic multi-classification for the data sets of the intervertebral foraminal stenosis.Finally,we utilized the transfer learning to optimize the overfitting problem of the deep learning algorithm in the small sample dataset.Experimental results show that the classification accuracy of this algorithm is 87.5% on the dataset of spinal foramina,and it has good robustness and generalization performance.(2)The simultaneous detection,segmentation and classification of spine structures play an important role in the early diagnosis of various diseases based on pathogenesis.When the lesion area and its adjacent structures are detected at the same time,it will provide more help for the radiologist to diagnose the disease according to the pathogenesis.According to the physiological structure of the human body,multiple structures of the spine are directly interdependent and interact with each other.In terms of multiple tasks under the framework of deep convolutional neural networks,different tasks can also affect each other.The multitask joint optimization in the spine overall mode is one of the solutions to seek the above-mentioned potential correlation dynamic balance.In this paper,we propose a novel end-to-end Multi-task Multi-structure Correlation Learning Network(MMCL-Net)for the detection,segmentation,and classification(normal,slight,marked,and severe)of three types of spine structure:disc,vertebra,and neural foramen simultaneously.And the model is locally optimized to achieve a more stable dynamic equilibrium state.Extensive experiments on T1/T2-weighted MR scans from 200 subjects demonstrate that MMCL-Net achieves high performance with m AP of 0.9187,the classification accuracy of 90.67%,and dice coefficient of 90.60%.(3)Precise segmentation for skin cancer lesions at different stages is conducive to early detection and further treatment.We propose a weakly supervised segmentation method for dermoscopy images through CNN responding superpixel regions(called CNN-SRR),given that the substantial cost of obtaining pixel-perfect annotation for these tasks.CNN-SRR combines a modified classifier based on deep learning and unsupervised superpixel algorithm.The former leverages abundant image-level labeled data to twist parameters to focalize in the lesion regions.The extraction of lesion region responses consists of two stages,which are training a CNN classifier and back-propagate the top layer peak of the classifier.Afterward,the image is over-segmented into a set of primitive superpixels that are merged into several regions as proposals,one of which is activated as the mask by lesion region responses via non-maximal suppression.Quantified experiments on ISBI2017 and PH2 datasets prove that the proposed method can effectively discriminate lesion regions and even achieve competitive accuracy to the supervised segmentation approaches.Specifically,the Jaccard coefficient and Accuracy of our method are improved by 12.4% and 3.3%,compared with the unsupervised superpixel segmentation algorithm,respectively.Conclusion: For the automatic multi-level task of intervertebral foramen stenosis,the multi-structure detection of spinal column in global mode,segmentation and lesion classification/grading tasks,and the segmentation task of skin cancer lesions under weakly supervised,this paper proposes three efficient and obvious advantages Deep learning algorithms are multi-level algorithm for intervertebral foramen stenosis(IFSNet),multi-structure and multi-task spine structure detection and segmentation algorithm(MMCL-Net)and CNN weakly supervised skin cancer image segmentation algorithm using superpixel region response(CNN-SRR).Among them,the successful construction of MMCL-Net provides an effective verification basis for deep learning algorithms to deeply mine the anatomical structural features of medical images.CNN-SRR algorithm completes the segmentation of skin cancer lesions under weak supervision for the first time,and provides the semantics-visual cues of segmentation can better assist doctors in understanding and interpreting algorithm prediction results.