Search for: All records

Reliable methods for tumor detection and brain abnormalities are crucial to help find diseases at early stages. Having accurate software that uses machine learning to identify abnormalities of the brain may prevent a disease progression if used on an MRI of the patient. In this paper, we develop a neural network to detect and highlight brain tumors present in MRI’s. Our model is designed to be more compact than typical CNNs to minimize prediction times while still maintaining prediction accuracy. The model uses a dice coefficient for the loss function as well as accuracy metric. We adopt Adadelta as the optimizer as it is more robust and eliminates the requirement of manually tuned learning rates. Our model reduces the prediction time with fewer layers and convolution filters, while allowing rapid convergence to a stable solution. In addition, the models hyper-parameters are being fine-tuned in an iterative process to ideally achieves better segmentation accuracy. Experimental results show that our model improves the performance compared with the state-of-the-art methods.