Dual-stage deep learning framework for brain tumor classification and localization using multimodal MRI scans

Brain tumor classification and localization are essential for accurate diagnosis and effective treatment planning. With the increasing use of MRI in clinical workflows, automated tools have become crucial for assisting radiologists in providing fast and reliable analysis. Traditional approaches for tumor classification and segmentation are often time-consuming and subjective, underscoring the need for deep learning-based frameworks that can enhance diagnostic accuracy and efficiency. This study used the TCGA-GBM dataset from The Cancer Imaging Archive (TCIA) to assess a dual-stage deep learning framework combining a classification model with a RESUNET for segmentation. TCGA-GBM, part of the broader TCGA project, focuses on Glioblastoma Multiforme (GBM), a highly aggressive brain cancer. The dataset includes 3,929 images, with 2,556 non-tumor (class 0) and 1,373 tumor (class 1) samples. The framework incorporated three convolutional neural network (CNN) architectures—MobileNet, NASNetMobile, and ResNet101—each enhanced with a Transfer Learning Layer for classification, followed by a RESUNET network for tumor localization. Transfer learning enabled the use of pre-trained weights, improving convergence speed and generalization. MobileNet offered a lightweight, efficient solution; NASNetMobile provided a strong balance between accuracy and computational cost; and ResNet101 delivered deeper feature extraction for higher precision. The RESUNET architecture, combining U-Net and residual learning, accurately segmented tumor regions, enabling effective integration of classification and localization within a unified framework. In the classification stage, the models achieved average accuracies of 0.9600 for MobileNet + Transfer Learning Layer, 0.9700 for NASNetMobile + Transfer Learning Layer, and 0.9500 for ResNet101 + Transfer Learning Layer, demonstrating their effectiveness in categorizing brain tumors. Performance was further evaluated using precision, recall, and F1 scores for both classes. MobileNet + Transfer Learning Layer achieved 0.95 precision, 0.99 recall, and 0.97 F1 for class 0, and 0.98, 0.92, and 0.95 for class 1. NASNetMobile + Transfer Learning Layer achieved 0.96, 0.99, and 0.98 for class 0, and 0.99, 0.93, and 0.96 for class 1. ResNet101 + Transfer Learning Layer achieved 0.97, 0.96, and 0.96 for class 0, and 0.93, 0.94, and 0.94 for class 1. For tumor localization, the RESUNET segmentation network accurately delineated tumor regions across all classification models. The proposed dual-stage deep learning framework effectively automates both classification and localization of brain tumors from MRI scans. The results demonstrate strong performance across all architectures, with NASNetMobile + RESUNET achieving the most balanced combination of precision and efficiency. Compared to baseline CNNs and prior hybrid models, our framework achieves superior class-wise performance and more accurate tumor localization, highlighting its methodological and practical advantages. This framework shows promise for integration into real-time clinical workflows. Future research will aim to expand the approach to additional tumor types and MRI modalities, optimize it for real-time deployment, and validate its generalizability across diverse clinical datasets. • A two-phase deep learning pipeline integrates transfer learning for binary tumor classification with RESUNet-based segmentation for detailed localization, improving both performance and computational efficiency. • Efficiently combines spatial and semantic information from multiple MRI sequences, enhancing detection of subtle tumor variations and improving boundary delineation. • The framework is adaptable to various backbone architectures and imaging protocols, facilitating seamless integration into diverse clinical environments.