The Effect of Input Size in Deep Learning Semantic Segmentation

Over the last few years, fully convolutional networks (FCN) methods have been broadly applied to image and video semantic segmentation tasks becoming the state of the art. In the medical area, FCNs have been evaluated for lesion and structures localization and segmentation. With the improvement of acquisition devices, the resolution and size of the images have increased. Usually, deep and large architectures produce the best results but requires large amount of computation for the training process. Amount of memory required to load large images can be a restriction. In this study, we investigated the impact of resizing images for training procedure. Generally, better results are obtained when the FCNs are trained with the original image size, however, our experiments showed that the same performance can be achieved by training the models with smaller images and increasing the size of the predicted mask, instead of inferring the mask by feeding larger images to the network.