%0 Generic
%A Albert, Steffen
%C Heidelberg
%D 2023
%F heidok:34188
%R 10.11588/heidok.00034188
%T Prediction of treatment response and outcome in locally advanced rectal cancer using radiomics
%U https://archiv.ub.uni-heidelberg.de/volltextserver/34188/
%X With the increasing number of medical images, deep learning is being used more and more in radiomics, but it suffers from small and heterogeneous datasets. To address this, a radiomics pipeline was developed for the prediction of the treatment outcome for neoadjuvant therapy in locally advanced rectal cancer (LARC), focusing on developing methods for dealing with small, heterogeneous multicenter datasets. For normalization, six different normalization methods (five statistical methods and one novel deep learning method) were investigated in multiple configurations: trained on all images, images from all centers except one, and images from a single center. The impact of normalization was evaluated in four tasks: tumor segmentation, prediction of treatment outcome, prediction of sex and prediction of age. For segmentation, there were only significant differences when training on one center, with the deep learning method being the best with a DSC of 0.50 ± 0.01. For the prediction of sex and treatment outcomes, the percentile method combined with histogram matching works best in all scenarios. The classification performance was evaluated using a published neural network. This network consists of two U-Nets sharing their weights, with segmentation as an additional task. The maximum AUC was 0.75 (95 % CI: 0.52 to 0.92) on the validation set. This is better than chance, but not better than using a classifier trained on clinical characteristics. In summary, normalization did help with the generalizability of the neural networks, but there is a limit to what can be corrected.