Published Paper Details:

Brain tumor diagnosis using magnetic resonance imaging (MRI) remains a challenging clinical task due to heterogeneous tumor appearance, overlapping intensity distributions, and significant interpatient variability. Accurate tumor segmentation and classification are essential for effective treatment planning and prognosis; however, manual analysis is time-consuming and subject to observer bias. This work presents an enhanced deep learning framework that integrates advanced preprocessing, an improved UNet-based segmentation network with Squeeze-and-Excitation (SE) blocks and Gaussian Error Linear Unit (GeLU) activation, and a comparative evaluation of three deep convolutional neural network architectures--ResNet50, ResNet101, and DenseNet121--for multi-class brain tumor classification. The segmentation module provides precise tumor localization, enabling the classification models to focus on clinically relevant regions and suppress background interference. The framework is evaluated on the Figshare brain tumor MRI dataset containing glioma, meningioma, and pituitary tumor classes. Experimental results demonstrate strong segmentation performance with a Dice coefficient of 91.8% and classification accuracy of up to 98.04% on unseen test data. Comparative analysis indicates that ResNet50 achieves the best balance between accuracy, generalization capability, and computational efficiency among the evaluated models. These results highlight the effectiveness of enhanced segmentation guided deep learning for reliable and real-time brain tumor diagnosis.