Segmentation And Detection Of Liver And Liver Tumor Based On Deep Learning

The application of deep learning in medical imaging is becoming more and more widespread.The main applications of deep learning are classification,segmentation and detection of medical images.This article details and summarizes the applications of deep learning for different organs and proposes the application of deep learning methods to liver tumor segmentation,which is mainly based on segmentation and detection in deep learning.The most important step in the clinical diagnosis and treatment of colorectal cancer liver metastases(CRLM)is the evaluation of liver cancer.This paper proposed a network model applied to CRLM,which can automatically complete the task of segmentation and detection of patients with abdominal liver tumors.The network model uses an improved cascaded network to assist physicians in subsequent diagnosis and quantitative analysis.The improved cascaded network is composed of U-Net and Mask R-CNN,which are used to segment and detect liver tumors,respectively.Step 1: The trained U-Net model is used as the first layer of the cascaded network to segment liver organs as regions of interest(ROI);step 2: the morphological active contour is clipped and extracted for the ROI region;step 3: The second trained U-Net model and Mask R-CNN model as the second layer of the cascaded network respectively complete the accurate segmentation and detection of liver tumors in the ROI.A large number of experiments have proved that the average Dice coefficient of liver metastasis segmentation of the cascaded U-Net model is about 74%,and the average mean accuracy(mAP)is 83%;the Dice coefficient of the tumor instance segmentation of Mask R-CNN is 67%(Confidence is 95%),and the mean Average Precision(mAP)is 88%.Through further experiments,this paper compares abdominal enhanced CT at different periods and proves the feasibility of the model.This paper proposes an optimized and improved cascaded network method,which is divided into data level and network model level.At the data level,affine transformation and PG-GAN(Progressive of GAN)are used for data augmentation.At the network model level,ResNet,SENet,and DenseNet are used to optimize the U-Net at the second layer of the improved cascaded network,which are named ResU-Net,RSU-Net,and RSDU-Net.ResU-Net improves the short-range connection process of network parameter transfer and replaces the original parameter transfer with the residuals of the parameters to optimize the network model.RSU-Net implements a feature rematching strategy by introducing a channel attention mechanism and assigning different weights to different channels of features.RSDU-Net optimizes the network model by applying DenseNet’s feature sharing and high parameter utilization to complete the long-range connection of network parameter sharing.For enhanced venous liver tumor data,ResU-Net,RSU-Net,and RSDU-Net respectively obtained Dice coefficients of 74.6%,75.8%,and 76.2%,and the correct detection rates were 84.66%,85.30%,and 85.65%..This proves that the optimization of the network model has good experimental results.