[Paper Review] GPU-Based Fuzzy C-Means Clustering Algorithm for Image Segmentation

Mishal Almazrooie, Mogana Vadiveloo+1 more|arXiv (Cornell University)|Jan 1, 2016

Medical Image Segmentation Techniques10 references3 citations

TL;DR

This paper proposes a GPU-accelerated Fuzzy C-Means (FCM) clustering algorithm for medical image segmentation, leveraging data parallelism on CUDA-enabled GPUs to achieve high-speed processing. By mapping each pixel to a single GPU thread, the method reduces execution time from 519 seconds (sequential) to just 2.33 seconds for a 1MB brain MRI dataset, achieving a 245-fold speedup while maintaining segmentation accuracy comparable to the sequential FCM.

ABSTRACT

In this paper, a fast and practical GPU-based implementation of Fuzzy C-Means(FCM) clustering algorithm for image segmentation is proposed. First, an extensive analysis is conducted to study the dependency among the image pixels in the algorithm for parallelization. The proposed GPU-based FCM has been tested on digital brain simulated dataset to segment white matter(WM), gray matter(GM) and cerebrospinal fluid (CSF) soft tissue regions. The execution time of the sequential FCM is 519 seconds for an image dataset with the size of 1MB. While the proposed GPU-based FCM requires only 2.33 seconds for the similar size of image dataset. An estimated 245-fold speedup is measured for the data size of 40 KB on a CUDA device that has 448 processors.

Motivation & Objective

• To address the computational inefficiency of sequential Fuzzy C-Means (FCM) clustering in processing large medical image datasets.
• To explore the feasibility and performance gains of parallelizing FCM using General-Purpose computing on Graphics Processing Units (GPGPU).
• To develop a scalable, data-parallel GPU implementation of FCM that maintains segmentation accuracy while drastically reducing execution time.
• To evaluate the proposed method using both qualitative and quantitative metrics, including Dice Similarity Coefficient (DSC), on a simulated brain tissue phantom dataset.

Proposed method

• The proposed method maps each image pixel to a single CUDA thread, enabling fine-grained data parallelism across 448 GPU processors on a Tesla C2050 device.
• The algorithm implements the standard FCM objective function (Equation 1) with iterative updates of cluster centers (Equation 3) and membership values (Equation 4) in parallel.
• Convergence is determined when the change in membership values across iterations falls below a threshold ε = 0.005.
• The implementation uses CUDA's kernel launch model to execute membership and cluster center update computations in parallel on the GPU.
• Execution time is measured using gettimeofday() and cudaEventRecord() for accuracy and verification.
• Segmentation accuracy is evaluated using the Dice Similarity Coefficient (DSC), comparing results against a ground truth dataset.

Experimental results

Research questions

• RQ1 Can a GPU-based implementation of FCM achieve significant speedup over sequential FCM for large medical image datasets?
• RQ2 How does the parallel FCM implementation preserve segmentation accuracy compared to the sequential version?
• RQ3 What is the scalability of the GPU-based FCM across different image sizes, from 20KB to 1MB?
• RQ4 How does the performance of the proposed GPU-FCM compare to sequential FCM in terms of execution time and speedup?

Key findings

• The proposed GPU-based FCM achieved a 245-fold speedup over the sequential FCM implementation when processing a 1MB image dataset.
• For a 40KB dataset, the speedup reached 245-fold, decreasing to 169-fold at 80KB, and maintaining 186–238-fold speedup for datasets larger than 300KB.
• The execution time of the proposed method was 2.33 seconds for a 1MB dataset, compared to 519 seconds for the sequential FCM2 implementation.
• The Dice Similarity Coefficient (DSC) results showed that the segmentation accuracy of the parallel FCM was statistically indistinguishable from the sequential FCM across all tested slices (91st, 96th, 101st, 111th).
• The method demonstrated strong scalability, with execution time increasing linearly with image size while maintaining high performance on larger datasets.