Evaluating ResNet Acrchtectures for Brain Tumor Classification Based on MRI Images

EasyChair Preprint 13622

Abstract

In the rapidly advancing field of medical image analysis, the selection of an efficient and accurate model is critical for timely and reliable diagnosis. This study aimed to evaluate the performance of different deep learning architectures by comparing three variants of the Residual Network (ResNet) model—ResNet-50, ResNet-101, and ResNet-152—in the classification of brain tumors using MRI images. Each model was assessed on key performance metrics including validation accuracy, precision, sensitivity, and F1-Score. The results revealed that ResNet-50 and ResNet-101 both achieved a score of 0.95, outperforming ResNet-152, which scored 0.93. Although all models showed similar average training losses, an increase in validation loss with model depth suggested a decline in generalization capability for ResNet-152. Furthermore, the analysis documented a rise in training time with the complexity of the model, highlighting the greater computational requirements of the more sophisticated architectures. ResNet-50 was identified as the optimal model due to its balance between accuracy and computational efficiency, making it the preferred choice for the classification of medical images when resources and time are limited. In summary, while all tested models displayed high performance, ResNet-50 offers the best combination of accuracy and efficiency, proving to be the most practical model for medical image classification in resource-constrained settings.

Keyphrases: Brain tumors, CNN, Classification, ML, MRI, ResNet, TL, Transfer Learning, machine learning

@booklet{EasyChair:13622,
  author    = {Szymon Kruszyk and Damian Ledziński and Sandra Śmigiel},
  title     = {Evaluating ResNet Acrchtectures for Brain Tumor Classification Based on MRI Images},
  howpublished = {EasyChair Preprint 13622},
  year      = {EasyChair, 2024}}