Convolutional Neural Networks For Biomedical Image Processing

In recent years,convolutional neural networks(CNN)have made remarkable achievements in many fields of natural image understanding.With the deepening of its theory and application,convolutional neural networks have been used for the biomedical image processing and they have arisen much attention from researchers.Convolutional neural networks can learn features representing the intrinsic properties of images and these features can be better used for classification or reconstruction after the non-linear mapping,and then for multiple biomedical image processing tasks.In consideration of the success in natural images,we have applied it into different biomedical research fields including deconvolution for the fluorescence microscopy,statistical iterative CBCT reconstruction and PET image segmentation,respectively.Deconvolution for fluorescence microscope image refers to an image processing method with the ability to reduce the degree of image blur,noise levels and imaging distortion,which is applied to enhance the resolution of the fluorescence microscopy.The CNN-based deconvolution model achieves an optimal estimation of the original image by learning the degradation process between the training image pairs.Compared with the classical deconvolution methods,this method can not only reduce the noises,suppress the artifacts and preserve the detail information,but also has strong generalization capability.What's more,the proposed method well performs when applied on the real datasets.CBCT reconstruction refers to the technology to reconstruct the patients' cross-sectional images using projection data of human body irradiated by X-rays,thus assisting the doctor with clinical diagnosis and treatment.CNN-based statistical iterative CBCT reconstruction algorithm is designed to alleviate the blurry of the images when using the classical methods based on the Hessian regularization term,which helps further improve the image reconstruction quality.The proposed algorithm can suppress the noises,eliminate the step effect and enhance the resolution of the images.PET image segmentation is used to determine the size and location of the tumor for clinical diagnosis and treatment,and has important significance in clinical practice.Though the softmax classifier can predict the classification result of pixels well with the textural feature of PET images,relative high proportion of misclassification exists.The CNN-based segmentation method using multiple features containing high-level textural features and original gray-level features of PET images has an outstanding performance,and the segmentation results have high accuracy and robustness.In conclusion,the deep convolutional neural networks can facilitate the development of biomedical image processing and deal with different biomedical image processing tasks,and thus it is important for the diagnosis and treatment of clinical medicine.