Detecting brain tumors using deep learning convolutional neural network with transfer learning approach

Accurate classification of brain tumor subtypes is important for prognosis and treatment. In this study, we optimized and applied non-deep learning methods based on hand-crafted features and deep learning methods based on transfer learning using softmax as classification and KNN and SVM as classification for features extracted from deep features of ResNet101. For non-deep learning techniques, we extracted multimodal features as input to machine learning classifiers. For convolutional neural networks, we optimized and applied GoogleNet and ResNet101with transfer learning approach. The performance was evaluated in terms of sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), false positive rate (FPR), total accuracy (TA), and area under the receiver operating curve (AUC) using Jack-knife 10-fold cross validation (CV) for the testing and validation of the dataset. For two-class classification, entropy features using SVM Gaussian yielded the highest performance with 93.84% TA and 0.9874 AUC, and GoogleNet yielded 99.33% TA. For Multiclass classification, the highest performance to detect pituitary tumor yielded 95.65% accuracy and 0.95 AUC using ResNet101 with transfer learning. Deep features from ResNet101 using KNN improved detection of pituitary tumor (98.80% accuracy, 0.99 AUC), glioma (93.47% accuracy, 0.93 AUC), and meningioma (93.36% accuracy, 0.89 AUC). The deep features ResNet101-SVM to detect pituitary tumor yielded performance (98.69% accuracy, 0.98 AUC). Deep learning methods with transfer learning along with softmax and KNN and SVM as classification outperformed traditional machine learning methods. This approach may prove useful for prognosis and treatment planning to achieve better clinical outcomes.

Automated summary

Accurate classification of brain tumor subtypes is crucial for prognosis and treatment. This study shows that deep learning methods with transfer learning outperform traditional machine learning methods, indicating potential for better clinical outcomes.