Deep learning segmentation of brain ischemic lesion from magnetic resonance images for three-dimensional modelling

Automated segmentation is important for early detection and treatments to reduce disability and death risks among brain stroke patients. The existing segmentation algorithm is limited due to its computationally expensiveness in achieving a small accuracy. This work aims to develop a computationally economical automated brain infarct segmentation from T1-weighted Magnetic Resonance Imaging (MRI) using convolutional neural network architecture U-Net, but with reasonable accuracy compared to existing algorithm. The data used is taken from the Anatomical Tracing of Lesion After Stroke (ATLAS) open-source dataset, consisting of 304 brain t1-weigthed MRI images. The data is divided into training, test, and validation sets according to the 8:1:1 ratio. The data is then pre-processed so that all of them have similar size for the U-Net input. Then, the U-Net architecture is generated using encoder depth of 7. Certain hyperparameters including the number of epochs, encoder depth, and optimizers are varied. The U-Net with encoder depth 7 and using Adam optimizer gives the highest accuracy and loss, which are 92.33% and 0.9771, respectively. Further comparison with previous works shows that the present U-Net beaten the regular U-Net and also gives relatively similar accuracy and loss. Future improvements on the present U-Net is necessary so that the accuracy can be increased further, computationally economic, and to produce a near accurate semantic segmentation of brain lesion.