A modified dual-task DenseNet121 with Grad-CAM interpretability for brain tumor classification from MRI

Abstract Diagnosing brain tumors from MRI scans is technically challenging but necessary. Early and correct characterization provides direct indications for subsequent treatment of and survival of the patient. While deep learning models show accuracy, most of them focus on just one classification task lacking practical tools for clinical use. We introduce Modified DenseNet121 (Dual-Task framework built on a modified DenseNet121 with a custom regularization stack. This is important to detect abnormalities (Normal vs. Abnormal) and identify three tumor types (Glioma, Meningioma, and Pituitary) in a single architecture. The model incorporates Gradient-weighted Class Activation Mapping (Grad-CAM) for qualitative spatial interpretability, and the model is set up within a PyQt5 desktop research prototype illustrating real-time dual prediction with typical inference latency of 80–130 ms (mean: 185.80 ± 402.83 ms; median: 127.99 ms; 95th percentile: 134.65 ms), probability visualization, Grad-CAM heatmap overlay, and feedback recording from the physician. Elaborate benchmarking experiments were performed against six competing CNN architectures DenseNet121 (baseline), ConvNeXtSmall, ConvNeXtTiny, EfficientNetB0, NASNetMobile and Xception over two large, publicly available MRI datasets: dataset 1, Kaggle, holding 7023 images; and dataset 2, Mendeley, having 12,064 images. Modified DenseNet121 achieved 99.85% binary accuracy and 99.00% multiclass accuracy on dataset 1, and 99.62% binary accuracy and 99.00% multiclass accuracy on dataset 2, achieving competitive performance with all benchmark architectures across both datasets. The zero-shot evaluation on an additional publicly available external MRI dataset achieved 99.75% binary accuracy and 99.20% multiclass accuracy, demonstrating transfer performance under an external evaluation setting. However, because dedicated image-level overlap analysis was not performed, these results are interpreted as evidence of external transfer performance rather than definitive evidence of complete dataset independence. The experimental results demonstrate high predictive performance across two large publicly available datasets and verify the technical feasibility of Modified DenseNet121 as a robust research prototype. This study offers reproducible, interpretable, and research prototype foundation for future clinical translation of neuro-oncological computer-aided diagnosis systems.