[Paper Review] A Time Series Graph Cut Image Segmentation Scheme for Liver Tumors

Laramie Paxton, Yufeng Cao+4 more|arXiv (Cornell University)|Sep 13, 2018

Radiomics and Machine Learning in Medical Imaging29 references3 citations

TL;DR

This paper proposes a semi-automatic graph cut method for liver tumor segmentation in 2D CT scans using time-series intensity features and normalized energy terms to eliminate λ-optimization. It achieves a mean Dice similarity coefficient of 0.77 and a processing time of 1.25 minutes per slice without requiring training data.

ABSTRACT

Tumor detection in biomedical imaging is a time-consuming process for medical professionals and is not without errors. Thus in recent decades, researchers have developed algorithmic techniques for image processing using a wide variety of mathematical methods, such as statistical modeling, variational techniques, and machine learning. In this paper, we propose a semi-automatic method for liver segmentation of 2D CT scans into three labels denoting healthy, vessel, or tumor tissue based on graph cuts. First, we create a feature vector for each pixel in a novel way that consists of the 59 intensity values in the time series data and propose a simplified perimeter cost term in the energy functional. We normalize the data and perimeter terms in the functional to expedite the graph cut without having to optimize the scaling parameter $λ$. In place of a training process, predetermined tissue means are computed based on sample regions identified by expert radiologists. The proposed method also has the advantage of being relatively simple to implement computationally. It was evaluated against the ground truth on a clinical CT dataset of 10 tumors and yielded segmentations with a mean Dice similarity coefficient (DSC) of .77 and mean volume overlap error (VOE) of 36.7%. The average processing time was 1.25 minutes per slice.

Motivation & Objective

• Address the challenge of low-contrast, poorly bounded liver tumors in CT imaging, which hinder accurate manual segmentation by radiologists.
• Reduce the time and variability associated with manual segmentation while improving reproducibility.
• Develop a computationally efficient, semi-automatic method that avoids training data requirements common in machine learning approaches.
• Improve segmentation robustness in the presence of noise and weak boundaries using time-series intensity data.
• Enable fast, accurate, and reproducible segmentation suitable for clinical use in pre-surgical planning and tumor staging.

Proposed method

• Construct a 59-dimensional feature vector per pixel using intensity values across 59 phases of dynamic CT scans to enhance tissue discrimination.
• Formulate an energy functional with a data fidelity term based on predefined tissue means from expert-identified ROI regions and a simplified perimeter cost term.
• Normalize both data and perimeter terms in the energy functional to eliminate the need for λ parameter optimization.
• Apply the Boykov-Kolmogorov max-flow algorithm to perform graph cuts in two stages: first separating tumor from healthy tissue, then vessel from tumor tissue.
• Use a two-step segmentation pipeline: initial cut separates tumor from non-tumor, followed by a second cut to isolate vessel structures.
• Leverage the max-flow min-cut theorem to compute global energy minimization efficiently, avoiding local minima issues common in active contour methods.

Experimental results

Research questions

• RQ1 Can incorporating time-series intensity data into a feature vector improve the robustness of graph cut segmentation for liver tumors with weak boundaries and low contrast?
• RQ2 Does normalizing the data and perimeter terms in the energy functional eliminate the need for λ parameter tuning while maintaining segmentation accuracy?
• RQ3 Can a semi-automatic method based on expert-identified tissue means achieve competitive performance without requiring large-scale training data?
• RQ4 How does the proposed method compare in accuracy and speed to existing graph cut and machine learning-based segmentation techniques for liver tumors?
• RQ5 What are the primary limitations of the method in detecting small tumors and accurately segmenting vessel structures?

Key findings

• The proposed method achieved a mean Dice similarity coefficient (DSC) of 0.77 on a clinical dataset of 10 liver tumors, indicating strong overall segmentation performance.
• The mean volume overlap error (VOE) was 36.7%, which is relatively high but comparable to other graph cut-based methods, especially considering the absence of training.
• The average processing time was 1.25 minutes per slice, significantly faster than the average 4.2 minutes per tumor reported for manual segmentation.
• The method outperformed five of nine other methods in runtime and achieved a DSC score comparable to manual segmentation (0.83) on the MIDAS dataset.
• The method showed a relatively high relative volume difference (RVD) of 35.1%, primarily due to over-segmentation and missed small tumors, indicating challenges with small or low-contrast lesions.
• The second graph cut stage performed poorly in vessel segmentation, suggesting a need for improved modeling or preprocessing for vascular structures.