I. Introduction
Colorectal cancer (CRC) is the third tumor in terms of incidence and mortality, and 60% of CRC are diagnosed as Locally Advanced Rectal Cancer (LARC)[1]. The recommended treatment is neoadjuvant chemoradiotherapy (nCRT) followed by Total Mesorectal Excision (TME)[2]. Despite the advantages shown by nCRT, patients' response varies widely, ranging from completely response (up to 20% of cases) to no response or tumor progression [3]. Artificial intelligence (AI) has shown promises in the development of radiomics signature, based on Magnetic Resonance Imaging (MRI), that can predict patient's response to nCRT, thus allowing more personalized treatments [4]–[7]. Despite the promising results, the translation of this approaches into clinical practice is still limited by many reasons, including the lack of automatic segmentation methods. Indeed, both manual and semi-automatic segmentations methods have two main drawbacks: they are a time-consuming task, that has to be regarded as prohibitive when very large databases are evaluated, and they lead to a high inter-reader variability that can strongly impact on the performance of predictive tools [6]. Therefore, developing automatic segmentation methods is of key importance to realize robust tools that can be effectively used in the clinical practice. In the last few years, Deep Learning (DL) algorithms have been used in the medical imaging field to segment and detect anatomical structures [8], [9]. More recently, the U-Net architecture [10] has been presented to overcome some limitations of previously developed structures, i.e., Fully Convolutional Networks (FCNs) and Convolutional Neural Networks (CNNs). The main advantage of the U-Net structure is the absence of the fully connected layer, replaced by the up-sampling layer and the deconvolutional layer, which allow to obtain a probability score map with the same size of the input, classifying each pixel instead the whole image [10]. To the best of our knowledge, only few studies used the U-Net to automatically localize and segment LARC on MR images [11]–[13]. However, all these methods require an initial manual crop of the image to delimit the region of interest. Moreover, none of them was validated on an external dataset.