I have a dataset of brain tumours images. and I have to build a model to classify the malignancy grade of these tumours. The size of the tumours varies from small to large. The ROI are already selected by an expert (the boundary of tumours is drawn).
To work on such dataset, for sigmintation step, do you recommend on cropping the tumours based on boundries, which will result in a variation of image dimensions as shown below. Or having the original image, but subtrackting the background of the tumor, which will result in fixed sized images of the tumor with black background, but has a large dimention considiering the tumor is only a part of the original image.
also, is it applicable working on cropped images that are resized to have same fixed sized images?