Image Segmentation Algorithms for Brain Tumor Detection. A Preliminary Analysis

This study presents a preliminary analysis of three image segmentation algorithms for brain tumor detection in Magnetic Resonance Imaging (MRI): Segment Anything Model (SAM), SAM for medical imaging, and a novel algorithm called "Brain Killer" (BK). The research utilized a public open-source dataset, focusing on meningioma, glioma, and pituitary tumors. SAM, a transformer-based model pre-trained on a massive dataset, showed high-quality results on instance segmentation task across all tumor types. SAM for medical imaging, optimized for DICOM files, showed improved precision in tumor boundary detection. Our algorithm, BK, a novel unsupervised algorithm based on patch-based k-Means clustering, provided detailed segmentation including internal tumor characteristics. The results underscore the complementary strengths of supervised and unsupervised approaches in medical image analysis, suggesting potential for integrated solutions in clinical applications.

I. Introduction

Brain tumors represent a significant challenge in the medical field, requiring accurate and timely diagnosis for effective treatment planning [1]. The advent of advanced imaging techniques, particularly Magnetic Resonance Imaging (MRI), has revolutionized the ability to visualize and characterize brain tumors. However, the interpretation of these complex images remains a time-consuming and expertise-dependent process. In recent years, the integration of Artificial Intelligence (AI) and computer vision techniques into medical image analysis has opened new avenues for enhancing the efficiency and accuracy of tumor detection and characterization.