Multiclass Brain Tumor Classification From MRI Scans Using Enhanced Inception Network With Residual Connections

MRI image classification of brain tumors into clinically significant categories—glioma, meningioma, pituitary tumor, and no tumor by employing deep learning models. The aim is to compare the performance of pre-trained and self-trained CNN for effective classification. The work is combining high-performance transfer learning models with an optimally designed training pipeline for brain MRI classification and accurate diagnosis is of prime importance. The models used are EfficientNetB0, InceptionV3, InceptionResNetV2, ResNet50, MobileNetV2, Darknet-53 variants, ResNet34, GoogLeNet (Inception v1), SqueezeNet, and ShuffleNet variants. The tested model got an accuracy reported by InceptionResNetV2 as 98.25, closely trailed by InceptionV3 at a test accuracy of 98.02% and ResNet50 at 97.03%. Pre-trained models are compared to other models such as SqueezeNet at 74.68% accuracy and GoogLeNet at 79.48%. Trained models such as Darknet-53 and ResNet34 did exceptionally but not beyond the top pre-trained models. Surprisingly, ShuffleNet achieves 84.97% is the best accuracy-efficiency trade-off.