Impact of network parameters on a U-Net based system for rectal cancer segmentation on MR images

The use of Deep Learning (DL) algorithms in the medical imaging field is increasing in recent years. However, they require the selection of a set of parameters to properly perform. In this study we evaluated the impact of three factors (the construction of the training set, the number of network layers and the loss function) on the performance of a U-Net system in the segmentation of Locally Advanced Rectal Cancer (LARC) on Magnetic Resonance Imaging (MRI). Images from 3 different institutions and 4 different scanners were used to this scope, for a total of 100 patients. All images underwent a pre-processing step to normalize and to highlight the tumoral area. The sequences of two scanners were used to construct the networks while the remaining sequences were employed for validating the best performing systems. From our results, it emerged that Dice Similarity Coefficient is not affected by any of the evaluated factors. Conversely, the choice of loss function could bias the results towards either precision or recall and, thus, it should be properly performed according to the scope of the network. Moreover, a slightly improvement of the performances was observed using a training set based on clustering, maybe due to a better representation of the heterogeneity characterizing medical images.