Evaluation Of Brain Tumor Models Using Optimal Deep Learning Approaches

Abstract

Brain tumor classification using Magnetic Resonance Imaging (MRI) is a pretty important task in medical image analysis, especially for helping early diagnosis and deciding treatment plans. Usually, doctors and radiologists do it by hand, manual interpretation, but that takes time and the results can vary a bit, sometimes quite a lot. In this paper, we introduce an explainable deep learning setup, built on the ConvNeXt-Tiny architecture, and we use transfer learning plus phased fine-tuning, for multiclass brain tumor classification. The method groups MRI scans into four broad labels: glioma, meningioma, pituitary tumor, and no-tumor. Before training, we do image preprocessing and use data augmentation to make the model more robust, and also to improve generalization. That includes normalization, rotation, flipping, zooming, and brightness adjustment. To make the whole system feel more transparent and dependable, Gradient-weighted Class Activation Mapping (Grad-CAM) is added, so we can see which tumor areas are actually driving the prediction, kind of like a visual clue. In the experiments, the proposed framework reaches 93.48% accuracy, 94.20% precision, 92.86% recall, 93.53% F1-score, and a 99.22% ROC-AUC score. Overall, the framework gives accurate results while staying interpretable, so it could fit well into intelligent clinical decision-support tools, where automated brain tumor diagnosis is expected to be both fast and understandable.