ResNet18 beats deeper models with 97% accuracy in brain tumor MRI detection

A team of researchers (Annapurna V K, Asha N, K Paramesha, Shabana Sultana, Kirankumar Humse) has published a paper on arXiv demonstrating an automated deep-learning approach for brain tumor detection from MRI scans. They employed Convolutional Neural Networks (CNNs) and Residual Networks (ResNet), specifically comparing ResNet18 and ResNet50 architectures. Using a dataset of 3,929 brain MRI images, they applied transfer learning to classify scans into tumor and non-tumor categories. Results show ResNet18 achieves 97% accuracy versus 96% for ResNet50, indicating that a shallower model generalizes better on limited medical data. The approach enables fast, accurate, and cost-effective detection without requiring massive computational infrastructure.

This work addresses a critical challenge in medical imaging: the complexity and variability of brain structures that make manual interpretation time-consuming and error-prone. By leveraging pretrained architectures, the method reduces the need for extensive labeled datasets and training time. The higher accuracy of ResNet18 suggests that overparameterized models may not be optimal for medical tasks with smaller datasets. The framework is designed to support early diagnosis and clinical decision-making, potentially reducing diagnosis delays and improving patient outcomes in resource-constrained settings. The paper is available on arXiv under ID 2606.27405.