Comparative Analysis of Transfer Learning Based Residual Networks for the Brain Tumours Detection.

Background: Brain tumours are critical disease and require an accurate and reliable method to detect the tumours. Therefore, many researchers are opting for deep learning models to classify brain tumours precisely for large and complex data sets. Purpose: This study explores the usage and performance of deep learning models, ResNet50 and ResNet101, for identifying multi-class brain tumours from 7,023 magnetic resonance imaging (MRI) sourced from Kaggle. Methods: Two models are proposed: ResNet50 based on transfer learning and fine-tuning, which outperform the proposed transfer learning-based ResNet101. Results: Comprehensive assessments are performed on these proposed models based on performance metrics having accuracy, precision, recall and F1 score, to ascertain the model's capability to highlight class discrimination. By focusing on these whole evaluation criteria, the work demonstrates that the proposed ResNet50 architecture, with an accuracy of 99.23%, can be achieved and effectively applied to automated brain tumour diagnosis from MRI and makes reliable decision making in healthcare demanding classification tasks and from this robust model performance and exact classification capabilities. Conclusion: ResNet50 with transfer learning and fine-tuning is a better option for applications that need to make decisions quickly with less GPU because it is more efficient in terms of processing, uses less training time, with better tumour detection and classification accuracy compared to ResNet101.