Liver Tumors And Pulmonary Nodules Detection In Medical CT Images Using Deep Learning

Liver cancer and lung cancer are both high morbidity diseases,and the fatality rate of them are increasing year by year.Artificial intelligence algorithm by deep learning is progressing continuously,which promotes the development of intelligent medicine rapidly.Deep learning algorithm plays an important role in the process of doctor's diagnosis and treatment of cancer.The automatic labeling of liver tumors and pulmonary nodules in medical CT images can be accomplished by constructing a detection algorithm based on deep learning,which assist doctors in diagnosis and treatment.Moreover,the automated deep learning detection algorithm can learn medical diagnosis and treatment knowledge through learning a large number of high-quality data,which are labeled by doctors.In this way,the integrated deep learning algorithm can be used in medical underdeveloped areas to help solve the problem of difficult medical treatment in underdeveloped areas.In this paper,the detection method of liver tumors and pulmonary nodules is mainly discussed.There are two methods,One is the target detection method to detect lung nodules,and the other is image segmentation algorithm to segment liver tumors.Which method to choose relys on the label of the data,if the label is location,the target detection method is used,if the label is the mask,the image segmentation algorithm is used.The main difficulty of lesion detection is that small lesions are difficult to detect,both small tumors and nodules are not easy to detect.In order to preprocess the medical images expediently,this paper proposes a medical CT images processing method to normalize different types of medical CT images.In order to solve the detection of small nodules,this paper presents a multi-scale lung nodule detection method based on dense connection network,which can detect nodules in two scales,especially,the network is more inclined to detect small nodules.Classification networks and an image block matching network are proposed to reduce false positives in detecting lesions.Before detection of liver tumors,organs need to be segmented.Unet network on segmenting liver increase a sequentially extracted model to reduce the amount of resources waste in engineering.Unet method based on feature fusion is used to detect multiscale tumors.This paper studies the advantages and disadvantages of morphological image processing method compared with segmentation algorithmic.Aiming to solve the problem of sample imbalance in the field of medical image processing,we study the impact of different loss functions on the network.