Degraded Medical Image Interpretation With Heuristic Deep Learning Methods

Recently,with the rise of big medical data and development of computer technology,the smart healthcare is becoming come true.And the artificial intelligence methods based on image processing and machine learning have become a new approach of medical image interpretation,which provide a new way for the diagnosis and treatment of cancer and other diseases.Deep learning is one of the most popular research fields of artificial intelligence at present.With powerful learning ability and generalization ability,it has made great success in natural image interpretation and even surpasses human.However,this latest method also has shortcomings including severe reliance on big data and poor interpretation on model effectiveness.In fact,the amount of medical image is not as abundant as the natural image since the difficulty of acquision,and the requirement of medical image interpretation is often much higher than natural images due to the low imaging quality.These specificities of medical images and the shortcomings of deep learning itself further limit the application of deep learning method in medical image interpretation.In order to further develop the advantages of deep learning and promote its application in medical image interpretation,this paper combines heuristic information in relevant fields with deep learning models to overcome the limitations and concentrates on the degraded medical image interpretation,which has mainly completed the following three tasks: 1.For the problem that noise model in megavoltage computed tomography(MVCT)images cannot be accurately estimated,a semi-supervised denoising method based on denoising autoencoder is proposed.Firstly,with the help of the existing medical image denoising method to obtain approximate high quality MVCT images as references,the original unsupervised blind denoising task is transformed into a semi-supervised way,which significantly reduces the difficulty of denoising.Secondly,the noises on medical images are modeled by the denoising autoencoder,combined with the global denoising and local enhancement strategies which considers the details of the region of interest.Finally,the proposed method is compared with another existing denoising method,and the experimental results show that the noise of MVCT images is modeled more accurately by the proposed method and the denoising performance is better.2.For the problem of severe noises and artifacts in low-dose CT(LDCT)images,an enhanced model based on gradient regularized convolutional neural network(GRCNN)is proposed.Compared to the traditional deep learning models for image restoration,we introduce a gradient loss term to the denoising regression paradigm,which is used to measure the preservation of the image details in the restoration process,so as to overcome the problem of blurring edges and loosing details caused by the previous deep learning models.Experimental results on the real clinical low-dose CT images show that our proposed method is superior to the existing state-of-the-art methods both on visual effects and evaluation criteria.Moreover,this method also has good generalization abilities,which can be further extended to other medical image enhancement tasks.3.For the problem of severe artifacts and low spatial resolution on low-field magnetic resonance imaging(MRI)images which makes it difficult for target segmentation,a sequential model of image enhancement and target segmentation is proposed.First,by using the proposed GRCNN model to enhance the low-field MRI image,the visibility of whole image and separability of the target area are improved.Then,combining the class imbalance learning idea with the single target segmentation of medical images,a cost sensitive weighted U-Net model is proposed to segment the stomach area in the MRI images.The proposed method is also compared with the traditional segmentation method Level Set,which is widely applied in the medical image field.The experimental results show that the method based on deep learning is obviously superior to the traditional method,and the application of class imbalance learning to the deep learning model U-Net boosts the performance of the segmentation.