I'm developing a U-net like model which segments the damaged tissue of the brain between two time-points in Multiple Sclerosis patients. The model is given the baseline and follow-up images as x and the segmentation mask as y. The images are 3D (192, 218, 192) and the model input size is (128, 128, 128) and is being trained by the dice loss. The model is based on this paper: https://www.sciencedirect.com/science/article/abs/pii/S0895611120300732
What I'm currently doing is a centre crop for each image before the training, and the training dice score seems to be learning properly, but in the validation data it is not.
I have read that random cropping 10 times or so helps to reduce overfitting, but I don't exactly know how to implement it. I have written this function to do the random cropping:
def random_crop(img_bl, img_fu, mask, width=128, height=128, depth=128): x = random.randint(0, img_bl.shape - width) y = random.randint(0, img_bl.shape - height) z = random.randint(0, img_bl.shape - depth) img_bl = img_bl[y:y + height, x:x + width, z:z + depth] img_fu = img_fu[y:y + height, x:x + width, z:z + depth] mask = mask[y:y + height, x:x + width, z:z + depth] return img_bl, img_fu, mask
Do I just apply this function 10 times before the training for each image? Or is there a way to include the random cropping inside the model and overlap the predictions of these 10 subvolumes?
Any tips, links to articles or insights are greatly appreciated. Thanks!