AUTOMATED EXTRACTION OF LIVER OUTLINES FROM COMPUTED TOMOGRAPHY SCAN IMAGES USING A CUDA-BASED SEGMENTATION METHOD

Y. Chen, Dengpan Li, Qingyuan Zhu, Chunjin Wang, Jonathan Li2019 · DOI

Abstract

Traditional medical image segmentation algorithms, such as the Fast Marching Method (FMM), Fuzzy Connectedness (kFOE), and 3D thinning, are computationally intensive and often struggle with the massive datasets generated by modern Computed Tomography (CT) scans. This paper presents an optimized pipeline that leverages the parallel processing power of Graphics Processing Units (GPUs) via NVIDIA’s CUDA platform. By redesigning these algorithms to minimize synchronization overhead and memory bottlenecks, the proposed method achieves significant speedups—up to 13.8x for vascular skeletonization—while maintaining the clinical accuracy required for surgical planning.

Introduction

The segmentation of the liver and its internal vascular structures is a cornerstone of modern diagnostic radiology and surgical preparation. However, the sheer volume of 3D CT data poses a significant challenge for traditional central processing unit (CPU) implementations. Sequential processing of these algorithms leads to long wait times, hindering real-time clinical workflows.

The authors identify three primary algorithmic bottlenecks:

  1. Fast Marching Method (FMM): Its reliance on a global min-heap data structure prevents efficient parallelization.
  2. Fuzzy Connectedness (kFOE): Parallel implementations often suffer from "edge competition" errors, where simultaneous thread updates lead to boundary inaccuracies.
  3. 3D Thinning: The iterative nature of extracting a one-pixel-wide skeleton from a 3D volume is extremely time-consuming on a CPU.

Methodology: A Parallel Pipeline

To address these challenges, the researchers developed a multi-stage GPU-accelerated pipeline. The process begins with hepatic segmentation, followed by vascular extraction, and concludes with thinning algorithm processing to extract the hepatic vessel skeleton.

The proposed image processing pipeline, from raw CT image to extracted hepatic vessel skeleton.

Algorithmic Improvements

  • Hepatic Segmentation (Improved FMM): The global min-heap was replaced with a "narrow band" approach. By processing pixels in independent thread blocks, the GPU significantly reduces the synchronization overhead that typically stalls parallel FMM.
  • Vascular Extraction (CUDA-kFOE): To overcome GPU memory constraints, a 256x256 lookup table for affinity values was introduced. Furthermore, a two-iteration correction scheme was implemented to resolve "edge competition" errors, ensuring that boundary voxels are accurately calculated across thread blocks.
  • Vascular Skeleton Line Extraction (12-Sub-Direction Thinning): This stage utilizes a kernel-based filtering method with 14 templates to identify deletable voxels. To maintain data consistency, a double-cache strategy was employed, allowing threads to read from one buffer while writing status updates to another.

Datasets and Experimental Setup

The study evaluated the performance of the CUDA-based methods using three distinct abdominal CT datasets. Each dataset consisted of 512 × 512 pixel images with varying slice counts (135 to 175 slices), representing typical clinical data volumes.

Sample CT scan images from the datasets used to evaluate the segmentation and extraction algorithms.

Results and Discussion

The GPU-accelerated approach demonstrated superior performance across all metrics compared to traditional CPU implementations:

  • Hepatic Segmentation: Achieved speedup ratios between 2.52x and 4.75x. The efficiency improved as the target liver area increased, indicating better GPU occupancy for larger workloads.
  • Vascular Skeletonization: This stage saw the most dramatic improvement, with speedups ranging from 11.6x to 13.8x. The total refinement time for the vascular skeleton dropped from approximately 27 seconds on a CPU to just 2.3 seconds on the GPU.

While the GPU results are nearly identical to CPU results, minor differences arise from the GPU's optimization for single-precision floating-point operations. However, these variances are negligible in a clinical context compared to the massive gains in processing speed.

Conclusion

This research demonstrates that porting traditional medical imaging algorithms to a parallel CUDA environment is both feasible and highly effective. By addressing specific parallelization bottlenecks—such as data structure dependencies and thread competition—the researchers have moved liver and vascular segmentation closer to real-time clinical application. Future work may involve automating seed point selection through deep learning and exploring hardware-agnostic frameworks like OpenCL to increase the accessibility of these high-performance tools.

Summary updated

Abstract

The traditional fast marching algorithm for segmentation of the liver is suitable for processing on the central processing unit (CPU) platform, however, it is not suitable for implementation on Graphics Processing Unit (GPU). The fuzzy connection algorithm is used to extract the blood vessels in the liver, but there is a calculation error. The refinement algorithm is very time consuming when extracting the target skeleton line from the 3D image. In this paper, the fast-marching algorithm and the thinning algorithm are improved, which can be applied to the GPU computing, The fuzzy algorithm is also improved, and the calculation error of the algorithm is solved, making it more suitable for medical image processing. The computing speed of GPU is far faster than CPU. Medical image processing is one of the earliest applications where the computing performance is improved by GPU. These three segmentation methods, fast marching method, fuzzy connecting method and refinement algorithm are very common in medical image segmentation. Because the increment of medical image data results in the extension of computing time for medical image processing, it is necessary to apply the high parallelism of the GPU to speed up these algorithms. The experiment results demonstrate the feasibility of our accelerating algorithm.
BibTeX
@inproceedings{chen2019automated,
  author = {Y. Chen and Dengpan Li and Qingyuan Zhu and Chunjin Wang and Jonathan Li},
  title = {{AUTOMATED EXTRACTION OF LIVER OUTLINES FROM COMPUTED TOMOGRAPHY SCAN IMAGES USING A CUDA-BASED SEGMENTATION METHOD}},
  booktitle = {˜The},
  year = {2019},
  doi = {10.5194/isprs-archives-xlii-2-w12-31-2019},
}
Figure