NeuroDEEP-CNN: A Multi-Task Attention Framework for Joint Classification, Localization, and Segmentation of Brain Tumors on MRI

Abstract

Brain tumors require rapid and accurate differential diagnosis across many radiological subtypes. Existing classifiers rarely combine classification, lesion localization, and segmentation in a single interpretable system. NeuroDEEP-CNN is a multi-task ConvNeXtTiny backbone with three decoder heads for classification, coordinate regression, and mask segmentation, fused through a hybrid coordinate-mask attention gate. The model was trained on 12,643 T1, T1C+, and T2 MRI images across 39 classes using a two-phase focal warm-up and progressive fine-tuning strategy with an 80/10/10 stratified split. At the best checkpoint (epoch 33, selected on validation accuracy), the model achieved 92.81% validation classification accuracy and 98.26% validation top-3 accuracy. The detection branch reached a validation mean distance error of 18.56 pixels, and the segmentation branch reached a validation Dice coefficient of 0.664. A five-method explainability suite confirmed predictions are grounded in pathological MRI regions. NeuroDEEP-CNN closely approaches the strongest of five widely used transfer-learning baselines on a related three-class benchmark, while covering thirteen times the number of classes and simultaneously providing validated tumor detection and segmentation outputs suitable for radiologist-facing deployment.