Improving brain tumor classification performance with an effective approach based on new deep learning model named 3ACL from 3D MRI data

Many machine learning-based studies have been carried out in the literature for the detection of brain tumors using MRI data and most of what has been done in the last 6 years is based on deep learning. Most of them have been designed to work with 2D data. Since many tumor-free slice images are in the models designed in 3D, the classification performance is less than in the 2D models. However, 2D models are unsuitable for practical applications as they use the slice image representing the best tumor image. Therefore, in this study, for brain tumor classification, a new 3 (Attention-Convolutional-LSTM) 3ACL deep learning model that will work with MRI data is presented. Attention, convolutional, and LSTM structures were designed in the same learning architecture in the 3 ACL models, which had an end-to-end learning strategy. Thus, the representation power of the features was increased. In addition, since the proposed model was designed in 3 dimensions, 3D MR images were used directly in the 3ACL model without transforming the 3D MR images into 2D data. Highly representative deep features are extracted from the fully connected layer of the 3ACL model. The feature set is passed to the SVM. Besides, the weighted majority vote technique, which used SVM prediction results conveyed from all slices, improved classification achievement. BRATS 2015 and 2018 datasets were used in this study. For the BRATS 2015 and 2018 datasets, the proposed approach gave 98.90% and 99.29% accuracies, respectively.

Automated summary

A new deep learning model is proposed for brain tumor classification using 3D MRI data. By designing attention, convolutional, and LSTM structures in the same learning architecture, the representation power of the features is increased. The model directly uses 3D MR images and achieves accuracies of 98.90% and 99.29% on the BRATS 2015 and 2018 datasets, respectively.