A comparative study of noise-affected magnetic resonance image segmentation using traditional and robust Otsu algorithms with region of interest determination

Abstract

Medical image segmentation systems face significant challenges when processing magnetic resonance imaging (MRI) images of brain tumors, especially given the low contrast and presence of noise, which affects the accuracy of tumor region identification. This study aims to compare the performance of the classical Otsu algorithm with Robust Otsu in segmenting noise-affected images, focusing on threshold stability and separation accuracy. The two methods were applied to MRI images from the BraTS 2023 database, which included 23 clinical cases, using the original images and corrupted copies with different noise levels (0, 0.03, 0.06, 0.1). The experiments were performed using MATLAB 2024, and the evaluation was based on quantitative indicators including Dice, IoU, MSE, and Accuracy. The results showed that the performance of classical Otsu deteriorated with increasing noise, while Robust Otsu maintained more stable performance, achieving higher Dice values, a significant decrease in MSE, and improved accuracy across all noise levels. These results confirm that developing the Otsu algorithm to be more robust is a practical, simple, and effective solution for improving the segmentation of blurred medical images without the need for high computational complexity or resource-intensive deep learning models.