Abstract: Accurate classification of brain tumors from medical images is essential for enabling timely diagnosis and effective treatment. This study aimed to develop an innovative method for the diagnosis of brain tumors through a Fuzzy Richards Functions-based Ensemble Network (FRE-Net). The parameters of the Richards function are optimized through Grid Search (GS) for selecting an optimal set of parameters. Our proposed method integrates three well-established pre-trained Convolutional Neural Networks (CNNs): MobileNetV1, MobileNetV2, ResNet50V2. To increase the robustness of these models, we incorporate a novel Lightweight Multiscale with Squeeze and Excitation (LiteMSSE) Block, which improves performance by enhancing multi-scale feature extraction and enabling the network to capture more detailed spatial information for focusing on the most relevant features to improve overall diagnostic performance. Additionally, probabilities from the individual models are aggregated using a Fuzzy Richards Functions approach, which reduces the error between observed and ground truth data, further enhancing detection accuracy. The key innovation of this study lies in the design of novel LiteMSSE Block and use of Fuzzy Richard Function, which together enhance multi-scale feature extraction and combines diverse model predictions intelligently. The proposed FRE-Net method achieves an impressive accuracy of 98.47% on the four-class Kaggle dataset and 99.00% on the BR35H dataset by highlighting its potential as a powerful tool for diagnosis of brain MRI more precisely. Through extensive evaluations, we determine that our proposed ensemble method outperforms individual backbone models and existing methods.

Key words: Deep learning')">Deep learning, Fuzzy richards functions, Ensemble network, Convolutional neural network, Brain tumorclassification