Deep Learning-Based Comparative Framework for Automated Brain Tumor Diagnosis Using MRI

Abstract

Brain tumors are one of the most dangerous medical conditions because they are tough to detect at an early stage and even harder to treat successfully once they progress, which is why they are often linked with high death rates. MRI scans have become the standard method for brain tumor diagnosis since they produce detailed and high-resolution images of the brain in a safe and non-invasive way. However, despite their usefulness, MRI scans are not always easy to interpret because reviewing hundreds of scans can be time-consuming, repetitive, and stressful for radiologists, and the results may vary depending on the doctor’s experience and judgment. With the growing use of Artificial Intelligence in healthcare, deep learning models, especially Convolutional Neural Networks (CNNs), have shown great promise in helping doctors quickly and accurately spot patterns in medical images that may be too subtle or time-intensive for humans to notice. In this research, we compared two robust CNN architectures, DenseNet201 and EfficientNetB3, both trained via transfer learning for brain MRI classification. Unlike earlier studies that often relied on smaller datasets with only around 3,000 images, we built our work on a much larger, balanced dataset of 8,000 MRI scans, evenly divided between tumor and non-tumor cases. This larger dataset helped the models learn more general and reliable features, which are crucial for medical applications where accuracy directly impacts patient care. Both models were fine-tuned and carefully evaluated using multiple performance metrics, including precision, recall, F1-score, and ROC-AUC, to assess not only how well they detected tumors but also how consistent and reliable their predictions were. The results showed that DenseNet201 achieved a validation accuracy of 96.85%, while EfficientNetB3 reached 96.48%, with both models also performing strongly across precision, recall, and F1-score. These results highlight that dataset size and model architecture both play a massive role in how accurate a deep learning system can be. While each model had its strengths, the overall takeaway is that deep learning can genuinely support medical experts by reducing errors, saving time, and improving consistency in brain tumor diagnosis. This technology has the potential to reduce radiologists' workload, improve the likelihood of detecting tumors earlier, and ultimately increase patient survival rates, making AI a strong candidate for integration into real-world clinical settings in the near future.