Enhancing Brain Tumour Segmentation in MRI Scans Using AI: A Comparative Study of U-Net and Transformer-Based Architectures

Waghmare, Shravani Ravindra (2025) Enhancing Brain Tumour Segmentation in MRI Scans Using AI: A Comparative Study of U-Net and Transformer-Based Architectures. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (1MB) | Preview
Preview	PDF (Configuration Manual) Download (334kB) | Preview

Abstract

Segmentation of brain tumours presents one of the crucial and complicated tasks of medical imaging since rapidly and precisely identifying this segmentation has a direct impact on the diagnosis and planning of treatment and more directly on the patient outcome. Radiologists have to manually segment MRI scans which is cumbersome, inefficient and subject to an individual decision making. The present project seeks to resolve these disadvantages by applying and contrasting two algorithms of artificial intelligence (AI) U-Net and Transformer-based architecture to automatically segment tumour areas on MRI brain image scans automatically. The paper starts with an open-source brain MRI data preprocessing, namely data normalization and data resizing. U-Net is trained and tested on measures of accuracy, Dice Score, and IoU. The model was highly precise with a low loss of 0.935 and an accuracy of 99.06 percent and a Dice Score of 0.974 showing high performance in segmentation. Additionally, to increase interpretability and reliability, an explainability component, GradCAM, was added that produces intuitively interpretable visual heatmaps of important tumour areas, almost similar to an X-ray. Moreover, the Transformer-based model is being developed to understand whether the attention-based architecture has additional improvements with the quality of segmentation. Besides providing a very high deep learning pipeline accuracy, this project also focuses on explainability and future flexibility in a clinical environment. The results help to generate reliable and intelligible AI systems in medical images analysis.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Zahoor, Sheresh UNSPECIFIED
Subjects:	Q Science > QH Natural history > QH301 Biology > Methods of research. Technique. Experimental biology > Data processing. Bioinformatics > Artificial intelligence Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Artificial intelligence Q Science > Life sciences > Medical sciences > Pathology > Tumors
Divisions:	School of Computing > Master of Science in Artificial Intelligence
Depositing User:	Ciara O'Brien
Date Deposited:	04 Jun 2026 15:26
Last Modified:	04 Jun 2026 15:26
URI:	https://norma.ncirl.ie/id/eprint/9345

Actions (login required)

View Item