An Optimized Cascaded CNN Approach for Feature Extraction From Brain MRIs for Tumor Classification

This study enhances brain tumor classification by leveraging pre-trained models and attention mechanisms, ultimately improving accuracy and reliability in medical imaging diagnostics through feature extraction. A generative adversarial network (GAN) is employed for artifact removal, and images are resized to 224 224 pixels. Four pre-trained models—VGG16, ResNet50, NASNet, and DenseNet121—capture distinct features, and a Custom-CNN integrates these models with the convolutional block attention module (CBAM). The bald eagle optimization (BEO) algorithm is used to optimize the hyperparameters while balancing exploration and exploitation. The proposed Custom-CNN outperformed all pre-trained models in classifying brain tumor data across three datasets (Br35H, FIGSHARE, and SARTAJ), achieving a highest accuracy of 0. 9722, sensitivity of 0. 9689, precision of 0. 9721, and an F-score of 0. 9708 for the Br35H dataset. The incorporation of CBAM (Custom-CNN-CBAM) resulted in consistent performance improvements, with accuracy increasing from 0. 9433 to 0. 9572 for the Br35H dataset and from 0. 9336 to 0. 9501 for the FIGSHARE dataset, achieving the highest F-score of 0. 9598. The enhanced version of the Custom-CNN-CBAM model utilizing the BEO, referred to as Custom-CNN-CBAM-BEO, demonstrated superior convergence speed and accuracy compared to models optimized with genetic algorithm (GA) and particle swarm optimization (PSO), highlighting the effectiveness of the BEO optimization approach. The evaluation of various Custom-CNN models incorporating attention mechanisms and optimization techniques revealed that the Custom-CNN-CBAM-BEO model achieved the best average rank of 5. 68 in the Friedman test, with the Bonferroni-Dunn test confirming its superiority.