Training on Polar Image Transformations Improves Biomedical Image Segmentation

A key step in medical image-based diagnosis is image segmentation. common use case for segmentation the identification of single structures an elliptical shape. Most organs like heart and kidneys fall into this category, as well skin lesions, polyps, other types abnormalities. Neural networks have dramatically improved results, but still require large amounts training data long times to converge. In paper, we propose a general way improve neural network performance efficiency on imaging tasks where goal segment roughly elliptically distributed object. We polar transformations original dataset, such that origin transformation center point This results reduction dimensionality separation localization tasks, allowing more easily Additionally, two different approaches obtaining optimal origin: (1) estimation via trained non-polar images (2) model predict origin. evaluate our method liver, polyp, lesion, epicardial adipose tissue show produces state-of-the-art polyp performs better than most architectures biomedical when used pre-processing step, generally improves across datasets architectures.