BrainDxNet: A custom convolutional neural network for brain tumor diagnosis with explainability

A brain tumor is the abnormal growth of cells within or around the brain.Tumors can vary in size, with some showing symptoms early and others growing significantly before detection due to slower symptom onset.Manual detection is a lengthy and time-consuming process that can introduce errors.Evaluating brain tumor detection systems using a single dataset may lead to biased or incorrect results due to the dataset's limited diversity.Additionally, existing models are not generalized on all datasets to detect brain tumors effectively.To overcome this limitation, it is crucial to use multiple datasets that encompass a wide range of tumor characteristics and patient demographics.To address these challenges, we propose a custom Convolutional Neural Networkbased Brain Diagnosis Network (BrainDxNet) model to identify brain tumors efficiently.By using diverse datasets, we can ensure accurate and reliable detection across various scenarios, leading to a more generalized and efficient system.Our research evaluates the proposed BrainDxNet model and several transfer learning models using a merged dataset developed by combining data from two Roboflow datasets, the Figshare dataset, and the Kaggle dataset.The proposed BrainDxNet model achieved an accuracy of 98.17% on the Roboflow combined dataset, 99.22% on the Kaggle dataset, and 95.27% on the Figshare dataset.To demonstrate the superiority of the proposed BrainDxNet model, we present a comparative analysis with existing methods.Additionally, for black-box model explainability, we employ CAM, Grad-CAM, and Grad-CAM++ to highlight the most informative regions in the images, helping the BrainDxNet model classify brain tumors more efficiently than others.Future research will focus on real-time and distributed model training processes to provide a dynamic brain tumor diagnosis strategy in the medical sector.