Multi-model of convolutional neural networks for brain tumor classification in magnetic resonance imaging images

Classification of brain tumors based on magnetic resonance imaging (MRI) images is often carried out using convolutional neural network (CNN). However, the classification performance still needs to be improved due to the varying sizes, shapes, and positions of tumors and complex brain structures. In this study, we proposed a multi-model of CNN for brain tumor classification based on brain MRI images. The multi-model of CNN involves several CNN models (Xception, DensNet-201, and EfficientNet-B3), which were constructed using the proposed algorithm. This algorithm works by combining the advantages of each CNN model using classification results rules, which are formed based on the highest and smallest accuracy and false positive values from training validation. The first model in the multi-model structure can be selected from the CNN model with the smallest or largest validation accuracy and connected to the CNN model with the lowest false positives. We used brain tumor MRI image datasets to evaluate the algorithm's performance, including the THOMAS dataset (Dataset 1) and the NICKPARVAR dataset (Dataset 2). The test results showed that the multi-model of CNN constructed with this algorithm produces the best accuracy of 97.74% for Dataset 1 and 99.69% for Dataset 2. From these results, the multi-model of CNN can outperform the single CNN model with an accuracy improvement of 1.29%-4.19% for Dataset 1 and 0.22%-0.61% for Dataset 2.