20–21 Nov 2026
Online Conference
Asia/Karachi timezone

Explainable Deep Learning for Brain Tumor Classification from MRI

21 Nov 2026, 13:10
5m
Online Conference

Online Conference

Online Conference

Speaker

Muhammad Waqas (Ghazali University, School of Technology, Karachi, Pakistan)

Discussion (methods + interpretation)

We used the public Brain Tumor MRI Dataset (Nickparvar, Kaggle; 7,023 images, four classes). Because the compilation repeats scans across its train and test folders, near-duplicate images were removed from training, and results were also checked on a leak-free test subset (1,179 images). A custom CNN was compared with ImageNet-pretrained ResNet50, EfficientNet-B0 and DenseNet121 using accuracy, precision, recall, F1 and AUC. Grad-CAM heatmaps were examined for correct and incorrect predictions. Differences between the top two models were within expected noise, and Grad-CAM is a qualitative, coarse signal; this is a research prototype, not a diagnostic tool.

Background

Brain tumor classification from magnetic resonance imaging (MRI) is an important task in medical imaging. Although deep learning models can classify MRI images, their decision-making processes are often difficult to interpret. High classification accuracy alone does not establish whether a model relies on relevant tumor regions or irrelevant image features, creating a need to evaluate model explanations alongside predictive performance.

Aim & Purpose

This study aims to compare a custom convolutional neural network with three pretrained architectures for classifying brain MRI images into glioma, meningioma, pituitary tumor, and no-tumor categories. Grad-CAM was used to visualize the image regions influencing model predictions and examined align with relevant tumor regions.

Result

Transfer learning clearly outperformed the custom CNN. On the held-out test set (1,600 images), ResNet50 and DenseNet121 reached 91.3% and 91.1% accuracy, compared with 85.8% for EfficientNet-B0 and 74.3% for the baseline CNN. Glioma was the weakest class (recall about 0.79), and 8-13% of tumor slices were misclassified as no tumor. Grad-CAM heatmaps generally focused on the lesion for correct predictions, while some errors showed attention drifting to the skull, image edges, or midline.

Email address 24G-BCS023@agu.edu.pk
Phone number +923708146824

Author

Muhammad Waqas (Ghazali University, School of Technology, Karachi, Pakistan)

Co-authors

Mr Anas Siddiqui (Ghazali University, School of Technology, Karachi, Pakistan) Mr Moosa Khan (Ghazali University, School of Technology, Karachi, Pakistan)

Presentation materials