A Novel EfficientNet-based Approach for Brain Tumor Detection

Abstract

Accurate and early detection of brain tumors is important to improve patient outcomes, but this has intrinsically been somewhat subjective and contemporarily time-consuming and expensive. We will outline an approach in this groundbreaker that empowered deep learning and the efficiency of EfficientNetV2 in revolutionizing brain tumor detection. Our approach outperforms existing methods on conventional brain tumor datasets, achieving more than 99.6% accuracy in the classification of tumors into clinically relevant subtypes. Only with an EfficientNetV2 as the backbone, combined with carefully designed data augmentation techniques, pruning techniques, and transfer learning strategies, the model can realize high accuracy with drastically fewer parameters and significantly less computation compared to other deep learning approaches applied for brain tumor detection. This makes our approach highly scalable and feasible for translation to resource-constrained clinical environments. We have also introduced a pruning technique, which makes it possible to reveal the insights into model compression and parameter efficiency. Thus, our work has huge implications for the potential that deep learning, and more so EfficientNetV2 architecture, holds in terms of changing the landscape of brain tumor detection to fast, accurate, and cheap diagnosis toward better patient outcomes and saving lives.