HAL is a multi-disciplinary open access archive for the deposit and dissemination of scientific research documents, whether they are published or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L'archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d'enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Most often, image segmentation is not fully automated and a user is required to lead the process in order to obtain correct results. In a medical context, segmentation can furnish a lot of information to surgeons, but this task is rarely executed. Artificial Intelligence (AI) is a powerful approach for devising a viable solution to fully automated treatment. In this paper, we have focused on kidneys deformed by nephroblastoma. Yet, a frequent medical constraint is encountered which is a lack of data with which to train our system. In function of this constraint, two AI approaches were used to segment these structures.First, a Case Based Reasoning (CBR) approach was defined which can enhance the growth of regions for segmentation of deformed kidneys with an adaptation phase to modify coordinates of recovered seeds. This CBR approach was confronted with manual region growing and a Convolutional Neural Network (CNN). The CBR system succeeded in performing the best segmentation for the kidney with a mean Dice of 0.83. Deep Learning was then examined as a possible solution, using the latest performing networks for image segmenta- tion. However, for relevant efficiency, this method requires a large data set. An option would be to manually segment only certain representative slices from a patient and, using them, to train a Convolutional Neural Network how to segment. In this article the authors propose an evaluation of a CNN for medical image segmentation following different training sets with avariable number of manual segmentations. To choose slices to train the CNN, an Overlearning Vector for Valid Sparsed SegmentatIONs (OV 2 ASSION) was used, with the notion of gap between two slices from the training set. This protocol made it possible to obtain reliable segmentations of tumor per patient with a low data set and to determine that only 26% of initial segmented slices are required to obtain a complete segmentation of a patient with a mean Dice of 0.897.