Date Approved

6-25-2026

Graduate Degree Type

Thesis

Degree Name

Applied Computer Science (M.S.)

Degree Program

School of Computing and Information Systems

First Advisor

Dr. Sorio Boit

Second Advisor

Dr. Tiehang Duan

Third Advisor

Dr. Rajvardhan Patil

Academic Year

2025/2026

Abstract

Brain tumor MRI classification is an important medical-imaging task because MRI scans contain complex anatomical patterns that can be time consuming to interpret manually. This study evaluates whether a pre-trained Vision Transformer can classify brain tumor MRI images consistently across datasets with different class structures. Three publicly available Kaggle datasets were used: Nickparvar, Br35H, and Figshare. Nickparvar and Figshare were treated as multi-class classification tasks, while Br35H was treated as a binary tumor/no-tumor task. Images were converted to three-channel format, resized to 384 × 384 pixels, normalized using ImageNet statistics, and augmented during training. The selected model was ViT-Base Patch 16, implemented in PyTorch using the timm library and fine-tuned separately on each dataset. The Vision Transformer achieved strong test performance across all three datasets. On Nickparvar, the model achieved 99.54% accuracy, 99.54% precision, 99.54% recall, 99.54% macro F1-score, and 99.82% ROC-AUC. On Br35H, it achieved 99.50% accuracy, 99.50% precision, 99.50% recall, 99.50% macro F1-score, and 99.69% ROC-AUC. On Figshare, it achieved 96.31% accuracy, 96.68% precision, 96.31% recall, 95.47% macro F1-score, and 99.42% ROC-AUC. Performance was lower on Figshare, likely because its three tumor classes have more similar visual and radiological characteristics and because conversion from MATLAB format to image files may have introduced preprocessing variation. Overall, the results suggest that a consistent ViT-based pipeline can achieve strong benchmark performance across independent brain tumor MRI datasets. The study is limited by its use of two-dimensional image slices, reliance on public dataset labels, and 5 absence of clinical validation; therefore, the findings should be interpreted as a benchmarking contribution rather than a clinical diagnostic system.

Download

Share

COinS