Liver tumor segmentation based on 3D convolutional neural network with dual scale

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Malignant liver tumors pose a significant health threat.
• Accurate segmentation of liver tumors in CT images is challenging due to low contrast and complex tumor characteristics.

Purpose of the Study:

• To develop an effective algorithm for liver and liver tumor segmentation in CT images.
• To address the difficulties of low contrast, blurred boundaries, and diverse tumor appearances in CT scans.

Main Methods:

• A novel three-dimensional dual path multiscale convolutional neural network (TDP-CNN) was designed for liver and liver tumor segmentation.
• A dual-path architecture was employed to balance segmentation performance and computational resources, with feature maps fused at the end.
• Conditional random fields (CRF) were utilized to refine segmentation results and improve accuracy.

Main Results:

• The TDP-CNN algorithm was evaluated on the public liver tumor segmentation (LiTS) dataset.
• For liver tumor segmentation, the algorithm achieved a Dice score of 0.689, Hausdorff distance of 7.69, and average distance of 1.07.
• For liver segmentation, the algorithm achieved a Dice score of 0.965, Hausdorff distance of 29.162, and average distance of 0.197.
• The proposed method ranked first in liver segmentation and second in liver tumor segmentation in the MICCAI 2017 competition.

Conclusions:

• The proposed TDP-CNN algorithm demonstrates robust performance for both liver and liver tumor segmentation.
• The method shows significant potential for clinical application in liver cancer diagnosis and treatment planning.