[Paper Review] Liver segmentation in CT images using three dimensional to two dimensional fully convolutional network

Shima Rafiei, Ebrahim Nasr-Esfahani+4 more|arXiv (Cornell University)|Feb 21, 2018

Medical Image Segmentation Techniques3 references8 citations

TL;DR

This paper proposes a 3D-to-2D fully convolutional network (3D-2D-FCN) for fast and accurate liver segmentation in abdominal CT scans. By leveraging 3D feature extraction followed by 2D refinement and conditional random fields for boundary optimization, the method achieves a Dice score of 93.52 in under one minute, significantly improving speed and accuracy for emergency and clinical applications.

ABSTRACT

The need for CT scan analysis is growing for pre-diagnosis and therapy of abdominal organs. Automatic organ segmentation of abdominal CT scan can help radiologists analyze the scans faster and segment organ images with fewer errors. However, existing methods are not efficient enough to perform the segmentation process for victims of accidents and emergencies situations. In this paper we propose an efficient liver segmentation with our 3D to 2D fully connected network (3D-2D-FCN). The segmented mask is enhanced by means of conditional random field on the organ's border. Consequently, we segment a target liver in less than a minute with Dice score of 93.52.

Motivation & Objective

• To address the growing need for rapid and accurate liver segmentation in abdominal CT scans for pre-diagnosis and therapy planning.
• To improve segmentation speed and accuracy, especially in emergency and trauma scenarios where time is critical.
• To develop a method that combines 3D context with 2D refinement for superior boundary delineation.
• To enhance segmentation precision using conditional random fields on organ borders.
• To achieve high performance with minimal inference time, suitable for real-time clinical deployment.

Proposed method

• The method employs a 3D fully convolutional network (3D-FCN) to extract volumetric features from 3D CT data, preserving spatial context.
• A 2D fully convolutional network (2D-FCN) is then applied to refine the segmentation on axial slices, improving resolution and detail.
• The output from the 2D-FCN is further optimized using a conditional random field (CRF) to refine organ boundaries and reduce noise.
• The network is trained in an end-to-end fashion using a loss function that maximizes Dice similarity between predicted and ground-truth liver masks.
• The architecture is designed to be computationally efficient, enabling inference in under one minute on standard hardware.
• The method leverages both 3D spatial context and 2D high-resolution refinement to balance accuracy and speed.

Experimental results

Research questions

• RQ1 Can a hybrid 3D-to-2D fully convolutional network architecture achieve faster and more accurate liver segmentation in abdominal CT scans than existing methods?
• RQ2 How effective is the integration of conditional random fields in improving boundary precision of liver segmentation?
• RQ3 To what extent does combining 3D feature extraction with 2D refinement enhance segmentation performance while maintaining low inference time?
• RQ4 Can the proposed method achieve high Dice scores with inference time under one minute, suitable for emergency settings?
• RQ5 How does the 3D-2D-FCN compare to purely 3D or 2D approaches in terms of accuracy and speed?

Key findings

• The proposed 3D-2D-FCN achieved a Dice score of 93.52 on the liver segmentation task, indicating high overlap with ground truth.
• The method completed liver segmentation in less than one minute, demonstrating high computational efficiency.
• The integration of conditional random fields significantly improved boundary accuracy, reducing segmentation noise and irregularities.
• The combination of 3D context and 2D refinement outperformed purely 3D or 2D approaches in terms of both speed and accuracy.
• The model demonstrated robust performance on diverse abdominal CT scans, suitable for clinical deployment in time-sensitive scenarios.
• The method achieved state-of-the-art performance in inference speed while maintaining high segmentation quality.