Deep Learning Based Liver Lesion Analysis On CT Images

Liver cancer is one of the most serious cancers in the world.It is of great benefit for patients to detect and treat liver cancer at an early stage.However in clinical practice,radiologists need to examine the abdominal CT images one by one to diagnose patients.This relies on the doctors’ experience and consumes a lot of time and manpower.In order to avoid time-consuming and heavy CT-reading work,automatic algorithms for liver and lesion analysis are in urgent need of research and development.When high-quality analysis is required,image segmentation can be used to obtain the area of the liver and lesion.While making simple diagnoses or using mobile devices,object detection is more suitable to determine the location of the liver and the lesions.Currently,problems of liver and lesion analysis methods exist as follows.2D models are used for liver lesion segmentation,where 3D information hidden in the CT image sequence is ignored.3D networks are too heavy to run on weak computers.Methods for object detection focus on accuracy instead of lightweight.In this paper,two liver lesion segmentation methods and a object detection method are proposed to tackle the above problems:· To explore the 3D information in CT images,a 3D high-resolution network named HR-Net3D is presented,which is composed of a series of subnetworks for feature extraction in multiple resolution and feature fusion modules.The proposed method is able to extract and fuse features from multi-resolution to get global contexts and local details.Exper-imental results show that HRNet3 D outperforms on the liver and lesion segmentation task.· To avoid huge computation caused by 3D models,a novel multi-view segmentation frame-work is proposed to segment the liver and the lesions in CT images simultaneously.A segmentation model called TowerNet and a novel joint loss function are also reported to detect tiny lesions.Experiments show that the proposed method works well on the liver and lesion segmentation task while the computation decreases.· For the problem that existing object detection methods are hard to run on low-performance devices,a feature extractor CSPTR and a lightweight object detection network YOLOv5TRs are proposed,where Transformer is applied.Experiments tell that CSPTR is good at learning global features,and that YOLOv5 TRs is accuracy and real-time,which is suit-able for mobile smart devices.