Multi-class segmentation model to identify 3 different abnormalities in the brain

Neural network to automatically segment tumor regions in brain of MRI images

• Used 484 training images from Decathlon 10 Challenge dataset and nibabel to extract the images and labels from the files.

• Generated "patches"of our data which are as sub-volumes of the whole MRI images in order to speed up training time and reduce the memory needed

• Standardized the values to have a mean of zero and standard deviation of 1 to reduce the range of MRI images

• Built a 3D U-net model that takes advantage of the volumetric shape of MR images to predict the regions affected by Edema,Non-enhancing and Enhancing tumor

• Used Dice Similarity Cofficient then Soft Dice as a loss function to face the heavy imbalance in segmentation and

• Convert prediction from probability into a category by using a threshold of 0.5

• Evaluated model's performance for by calculating the sensitivity, specificity