Comparative Evaluation of CNN and Transformer-Based Models for Brain Tumor MRI Segmentation

Abstract

Brain tumors represent a critical medical problem with significant impact on human health. Manual assessment processes are insufficient in addressing this problem due to the difficulty in identifying tumor regions and disagreements among experts. This has led to an increasing demand for automated tumor detection or segmentation using machine learning and deep learning methods. This study compares the segmentation performance of selected deep learning models to investigate which model architecture yields better results under specific conditions. Two main approaches are considered: CNN-based (FCN, U-Net, Attention U-Net, DeepLabV3+) and a Transformer-based model (Swin U-Net). To ensure a fair comparison, all models were trained on the same dataset (BRISC 2025) under identical training conditions. Model performance was evaluated using Dice, IoU, Precision, Recall, and HD95 metrics. Among the tested models, DeepLabV3+ achieved the best performance with a Dice score of 0.8575 and an HD95 value of 7.8367. The findings indicate that CNN-based models outperform the Transformer-based model under the given dataset and experimental conditions. The results suggest that the performance of Transformer-based models may be sensitive to dataset characteristics. This study contributes to the literature by systematically evaluating CNN- and Transformer-based models under identical experimental settings.