Ct segmentation of liver and tumors fused multi-scale features

Aihong Yu, Zhe Liu, Victor S. Sheng, Yuqing Song, Xuesheng Liu, Chongya Ma, Wenqiang Wang, Cong Ma

Research output: Contribution to journalArticlepeer-review

Abstract

Liver cancer is one of frequent causes of death from malignancy in the world. Owing to the outstanding advantages of computer-aided diagnosis and deep learning, fully automatic segmentation of computed tomography (CT) images turned into a research hotspot over the years. The liver has quite low contrast with the surrounding tissues, together with its lesion areas are thoroughly complex. To deal with these problems, we proposed effective methods for enhan-cing features and processed public datasets from Liver Tumor Segmentation Chal-lenge (LITS) for the verification. In this experiment, data pre-processing based on the image enhancement and noise reduction. This study redesigned the original UNet with two novel modules and named it DResUNet which was applied deformable convolution. The first module aimed to recalibrate information by the channel and spatial dimension. The other module enriched deep information of these liver CT images through fusing multi-scale features. Besides, we used cross-entropy loss function for adaptive weights to solve the troubles of class imbalance in the dataset samples. These can improve the performance of the network in-depth and breadth feature learning to deal with many complex segmentation scenes in abdominal CT images. More importantly, the effect of predicted images fully proved that our methods are highly competitive among the segmentation of liver and liver tumors.

Original languageEnglish
Pages (from-to)589-599
Number of pages11
JournalIntelligent Automation and Soft Computing
Volume30
Issue number2
DOIs
StatePublished - 2021

Keywords

  • Computed tomography (CT)
  • Deep learning
  • Fully automatic segmentation
  • Liver tumors
  • Multi-scale features

Fingerprint

Dive into the research topics of 'Ct segmentation of liver and tumors fused multi-scale features'. Together they form a unique fingerprint.

Cite this