Effects of Image Size on Deep Learning

This paper presents the evaluation of the effects of image size on deep learning performance via semantic segmentation of magnetic resonance heart images with U-net for fully automated quantification of myocardial infarction. Both non-extra pixel and extra pixel interpolation algorithms are used to change the size of images in datasets of interest. Extra class labels, in interpolated ground truth segmentation images, are removed using thresholding, median filtering, and subtraction strategies. Common class metrics are used to evaluate the quality of semantic segmentation with U-net against the ground truth segmentation while arbitrary threshold, comparison of the sums, and sums of differences between medical experts and fully automated results are options used to estimate the relationship between medical experts-based quantification and fully automated quantification results.