MR image normalization dilemma and the accuracy of brain tumor classification model

In clinical practice, a detailed medical history and physical exam, which includes a thorough neurological examination, are used to diagnose a brain tumor. The size, form, margin, and texture of the tumor, among other things, can influence the diagnosis of brain tumors. Even tumor types that are pathologically different could have a similar texture and appearance in radiology. Furthermore, carefully reviewing all test results could take a significant amount of time for doctors and radiologists. As a result, more advanced medical technology, such as an automated system to diagnose brain tumors, is required. This study aimed to use artificial neural networks to develop an automated approach for detecting brain tumors in magnetic resonance imaging (MRI) scans. Towards this, about 4314 MRI images were acquired in this study. The data contains four classes: normal healthy brain, brain images having glioma, meningioma, or pituitary tumor. The raw data undergoes several preprocessing steps, and the impact of each preprocessing stage on the model accuracy was evaluated. A Densely Connected Convolutional Network (DenseNet) was trained using three different datasets. Enhancing the MRI image's contrast and normalizing its intensities improve the classification accuracy. It shows that preprocessing steps improved the learning convergence of DenseNet training. The proposed model achieved an accuracy of 96.52%, and the sensitivity and specificity were 98.5% and 82.1%, respectively, using ten-fold cross-validation. Hence, we conclude that specific preprocessing steps significantly enhance the tumor segmentation performance for automated systems when using advanced techniques such as deep learning.