We evaluated the influence of normalization (setting mean and standard deviation, histogram matching and percentiles) on the segmentation of rectal cancer on multimodal images when operating on multicenter data as part of a Radiomics pipeline. We used two different networks for segmentation. When training and evaluating on all data or data from a single center, normalization did not play a significant role. In contrast, when training on one center and evaluating on all others, it did play a major role. Best results are obtained by normalization using percentiles. Fixing the mean and standard deviation did not work well.
This abstract and the presentation materials are available to members only; a login is required.