A Novel Framework for Classification of MRI Images to Diagnose Brain Tumors using DenseNet 201

Brain cancer tumors are often undetected by radiologists due to the lack of clarity in some MRI images, making early detection of a potentially malignant cancer very difficult. One way to address this issue is with the use of convolutional neural networks (CNNs) with segmentation and classification steps to detect the unseen tumors. These networks can be utilized by radiologists as a computer aided tool to improve proficiency and develop pattern recognition of biomarkers. CNNs take in MRI data inputs and separate each scan into edges, colors, and sizes to apply an algorithm for classification, minimizing the error value with each iteration. The computer vision methods of image preprocessing, image segmentation, classification and pattern detection allow the model to be a viable supplement to clinician based diagnosis. Currently, studies have used various algorithms to improve cancer detection, but the accuracy rates have not been high enough for proper detection. This research aimed to produce a network with a higher validation set accuracy rate, training on 25 epochs. The model was trained on 70% of the images in the selected dataset, 20% for validation, and 10% of the data for testing. This study utilizes DenseNet architecture and is able to reach an accuracy rate for the brain tumor images of about 97%, higher than many other models and fit for classification usage. A limitation of this study is that the selected data set is confined to glioma and meningioma cancer types, leaving out other potential cancer a patient may have. However, as models continue to develop, using CNN based detection will ultimately benefit patients with various types of cancer in finding their tumors earlier for better prognosis.