Classification of brain tumors with deep learning Determination of the effects of data augmentation and transfer learning on classification Determination of effect of MR planes on classification In this study, the Convolutional Neural Network (CNN) was used to classify three different brain tumors (glioma, meningioma and pituitary) from T1-weighted MR images. The initial weights of the CNN architecture were transferred from the DenseNet121 network. The features obtained from the CNN architecture were classified by support vector machine (SVM), k-nearest neighbour (kNN) and Bayes methods. The performance of the classifiers was determined by accuracy, sensitivity, specificity, area under curve, and Pearson correlation coefficient (R) on the test data.Figure A. Flowchart of the proposed methodsPurpose: In this study, the essential aims are the classification of brain tumors on the figshare dataset; determination of which axial, coronal, and sagittal MR plane slices is effective in classification; and measurement of the increase in classifier performance resulting from the data augmentation on the medical image.
Theory and Methods:In this study, brain tumors were classified with CNN architecture using 3064 slices of T1-weighted MR images obtained from 233 patients in the figshare data set. Axial, coronal and sagittal MR plane slices were separated. Images were augmented with affine transformation and pixel-level transformation techniques. In the study, the weights of DenseNet121 network developed for ImageNet were transferred to CNN architecture. Features from the first fully connected layer of trained CNN architecture were used as inputs for SVM, kNN and Bayesian classifiers, and the results of these classifiers were compared to CNN result.
Results:After data augmentation, the R values of the CNN architecture increased from 0.880 to 0.967 and the accuracy values increased from 0.946 to 0.986. Accuracy values of SVM, kNN and Bayes classifiers were calculated as 0.998, 0.991 and 0.893, and R values as 0.995, 0.989, 0.790, respectively.
Conclusion:According to the results obtained with test data, CNN based SVM method has achieved higher performance compared to the literature. Brain coronal plane is more effective in tumor classification