Data Preprocessing Influence on the Medical Images Segmentation Quality

The paper is related to automating computer-tomography image processing. The main task of which is the segmentation of abdominal organs. A typical segmentation pipeline includes the following three stages: image preprocessing, segmentation, and mask post-processing. This article is devoted to the first stage, namely, medical image preprocessing. Generally, research related to medical image segmentation has not paid enough attention to data preprocessing research; usually researchers use conventional preprocessing such as normalization without a deep understanding of the effect on quality. The research is based on the state-of-the-art UNERT neural network and a widely used data set MICCAI 2015 with 30 abdominal clinical CT scans. They consist of 13 labeled abdominal organs in Nifty data format. Different methods of image normalization and preprocessing were investigated from the physical meaning of the Hounsfield scale and from the final quality point of view.