Research On Deep Learning Based Medical Image Enhancement Algorithm

In recent years,with the continuous improvement of hardware computing power and popularization of deep learning algorithms,digital medical image processing technology has developed rapidly.Computer-aided diagnosis(CAD)has become increasingly mature,and the diagnostic accuracy in some disease fields has reached or exceeded the level of professional Radiologists.With the rapid development of technology,CAD systems can work long-lastingly and efficiently,freeing doctors from a lot of repetitive work.However,the current CAD algorithm optimization encounters bottlenecks in many tasks.This paper proposes to break through the bottleneck at the input level through the research of image enhancement algorithms.Due to the existing evaluation indicators of natural images such as PSNR,SSIM has some limitations that make it is not proper to fully evaluate the advantages and disadvantages of medical image processing algorithms.This paper introduces CAD system instead of doctors to provide subjective evaluation indicators for algorithms,and this semantic evaluation indicators can break through the limitations of traditional methods and open up new medical image enhancement algorithms direction.This paper takes a segmentation task for CTP modality,a high-dimensional medical image,and a detection task for CT modality image as the starting points,to study how to improve the performance of CAD system through effective medical image enhancement algorithm.The research content of this paper is as follows:First,for CTP modality high-dimensional data,this paper decouples the difficult problem of high-dimensional image segmentation into image enhancement based on time-dimension reduction and modality transformation,making full use of the idea of cross modality method,combined with the proposed high-dimensional reduction method in image's time-dimension.Splitting hard task into two sub-tasks greatly reduces the difficulty of learning for CAD systems.And the SegMaxout network structure is proposed to effectively suppress the segmentation false positive regions.Finally,a new combined loss function is proposed to balance the gradient ratio of the negative and positive areas during the training process and stabilize the training process.This method has achieved a significant improvement in the ISLES 2018 dataset compared to the existing method,and achieved the champion of this challenge in the Medical Summit MICCAI 2018.Secondly,for low-resolution CT medical images,this paper proposes a PN-sample balance medical image super-resolution training strategy,which effectively solves the problem that the imbalance in training phase caused by the existing training methods and the degradation of the performance of the CAD detector system.Furthermore,the 3D dense DenseSRNet based on spatial dimension and feature attention mechanism(Attention)is proposed,which achieves a good reconstruction performace in the LUNA dataset and improves the performance of the CAD detector system.