Enhanced U-Net Architecture for Accurate Brain Tumor Segmentation in MRI Scans

Abstract

Precision in medical image segmentation is vital for accurate diagnosis and treatment planning in the contemporary healthcare system. Deep learning techniques, such as Convolutional Neural Networks (CNNs), U-Net architectures (UNETs), and Transformers, have revolutionized this field by automating laborious manual segmentation processes that were previously performed manually. Nevertheless, challenges such as intricate structures and indistinct features persist, leading to accuracy concerns. Scientists are diligently striving to address these challenges to fully harness the promise of medical image segmentation in the healthcare revolution. The objective of our study is to introduce an improved version of the U-Net model that is specifically tailored for the segmentation of brain cancer MRI images. The primary aim of this enhancement is to boost the precision of the segmentation process. Our plan consists of three main components. Initially, our main focus is on enhancing features by employing techniques such as Contrast Limited Adaptive Histogram Equalisation (CLAHE) during the image preprocessing stage. Furthermore, we enhance the architecture of the U-Net model by prioritizing a tailored layered design to enhance the quality of segmentation results. Ultimately, we employ a Convolutional Neural Network (CNN) model for post-processing to enhance segmentation results by utilizing additional convolutional layers. Our model was tested, validated, and trained using a total of 3,064 brain MRI images, with 612 images used for testing, 612 images used for validation, and 1,840 images used for training. We achieved outstanding results in terms of recall (93.66%), accuracy (97.79%), F-score (93.15%), and precision (92.66%). The Dice coefficient's training and validation curves exhibited minimal variability, with training achieving approximately 93% and validation reaching 84%, indicating a strong capacity for generalization. The high accuracy of the segmentation findings was confirmed through visual assessment, however rare errors such as false positives were observed.