Effects of Cropping vs Resizing on the Performance of Brain Tumor Segmentation Models

This study investigated the effects of image resizing vs cropping on the performance of state-of-the-art models for the brain tumor segmentation task. This is particularly important since many studies simply resize the image without thinking about potential effects of image distortion on the model's segmentation performance. Since the objective of tumor segmentation is to predict the pixels that comprise the actual tumor, image cropping was performed in order to focus more on the brain and the tumor and not on the background. This conjecture was tested using state-of-the-art models namely 2D U-Net, 2D U- Net with VGG19 as backbone, 2D U-Net with InceptionV3 ad backbone, and 2D U-Net with InceptionResNetV2 as backbone. Three different configurations were designed for this purpose. The first configuration used resized images while the second configuration used cropped images. The third configuration used pretrained weights of models of trained on the resized images and then applied them on the cropped images. Overall, the top three models are 2D U-Net with InceptionResNetV2 as backbone trained using the resized images followed by 2D U-Net trained also using the resized images and then finally by 2D U-Net trained using the cropped images. As to why cropping did not perform well in this experiment, several plausible explanations were provided in this study.