Lightweight deep learning for medical imaging using MobileNetV2-based brain pathology classification with Grad-CAM interpretability.

Introduction: Computed Tomography (CT) brain scans are crucial for diagnosing various neurological conditions, including tumors, cancer, and aneurysms. CT brain scans are essential for guiding treatment decisions and monitoring disease progression. In this study, we propose a novel framework for brain CT image classification that leverages convolutional neural networks (CNNs), Digital Imaging and Communications in Medicine (DICOM) pre-processing, transfer learning with multiple deep models, and ensemble prediction techniques. The primary objective is to enhance the accuracy and interpretability of brain abnormality detection. Methods: Our approach uses a comprehensive dataset of CT brain scans that undergo meticulous pre-processing to ensure data integrity and uniformity, and employs Grad-CAM for interpretability. We employ four state-of-the-art pre-trained models: MobileNetV2, ResNet-50, EfficientNet-B0, and VGG-16, each serving as a feature extractor, followed by a classification head tailored to our specific task. Results: The experimental results demonstrate that MobileNetV2 and the Ensemble model achieved the highest classification accuracy of 97.44% with macro-AUC scores of 0.9895 and 0.9914, respectively, followed by VGG16 with 92.31% accuracy and the highest macro-AUC of 0.9962. In contrast, ResNet50 and EfficientNetB0 achieved accuracies of 61.54% and 33.33%, respectively, indicating fundamental limitations in learning discriminative features for this medical imaging task. MobileNetV2 proved to be the most efficient model, achieving superior accuracy with training and test times of 89 s and 38.16 s, respectively. Discussion: MobileNetV2 is highly suitable for clinical deployment where both accuracy and computational efficiency are critical.