Bacterial Foraging Optimization-boosted convolutional neural network for brain tumor detection using MRI images

Objectives:

Accurate brain tumor classification from magnetic resonance imaging (MRI) is important for early diagnosis, treatment planning, and clinical decision support. However, manual MRI interpretation is time-consuming, depends heavily on expert experience, and may suffer from inter-observer variability. Although deep learning and traditional machine learning methods have been widely explored for automated brain tumor detection, many existing approaches still face challenges related to feature redundancy, computational cost, and reliable classification using limited medical image datasets.

Methods:

This study proposes a hybrid deep learning-based feature engineering and traditional machine learning classification framework, referred to as a Stacked Network Model (SNM), for binary brain tumor classification using MRI images. A publicly available MRI dataset containing 1275 tumor-positive and 1255 tumor-negative images was used. After augmentation, the dataset included 1574 tumor-positive and 1572 tumor-negative images. Four pre-trained convolutional neural network models, namely ResNet50, Xception, InceptionV3, and DenseNet121, were used as deep feature extractors. To reduce feature redundancy and improve discriminative capability, Bacterial Foraging Optimization (BFO) was applied as a feature selection strategy. The optimized deep features were then classified using traditional machine learning classifiers, including Logistic Regression, Support Vector Classifier, Random Forest, Decision Tree, XGBoost, K-Nearest Neighbor, and Naive Bayes.

Results:

The proposed ResNet50-BFO-LR pipeline achieved the best performance, with 99.81% accuracy, precision, recall, and F1-score on the test split. BFO reduced the ResNet50 feature dimension from 2048 to 991 features while improving classification performance compared with the non-optimized feature setting. The final model also achieved an average inference time of 0.11 s per image, indicating its potential for fast computer-aided MRI screening. Compared with previously reported methods on related brain MRI classification tasks, the proposed framework shows competitive or higher performance while maintaining a simple and efficient classification pipeline.

Conclusion:

The findings suggest that combining pre-trained CNN feature extraction, BFO-based feature selection, and traditional machine learning classifiers can provide an effective and computationally efficient framework for MRI-based brain tumor classification. However, further validation on external and multicenter datasets is required before clinical deployment.