Extracted features from pre-trained networks for brain tumor classification: A performance and computational cost analysis of ML, ANN, and CNN

In this paper, a comparative analysis of different machine learning (ML) and deep learning (DL) models for brain tumor image classification is presented. Traditional ML algorithms — Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LR); Custom Artificial Neural Network (ANN), Convolutional Neural Network (CNN), along with VGG16 model for feature extraction coupled with ANN. AUC and execution time were used to evaluate the performance of all models based on classification accuracy. The custom CNN obtained the perfect classification accuracy in the brain tumor dataset; however, it sacrificed computational time. The combined VGG16+ANN, have an increase in execution time but there is a jump related to other models, like the next from the list of most accurate ensembles. Random Forest is also the lowest for AUC and SVM, Logistic Regression have average classification accuracy with fast training. Combined with ANN, the VGG16 model delivers computational overhead comparable to smaller models yet still provides competitive accuracy. In this work, we have examined trade-offs between computational efficiency and classification performance in the context of medical image analysis tasks. We expect our study to provide researchers and practitioners with useful information for choosing models for a variety of constrained settings related to the clinical deployment of medical imaging classifiers.