The analysis of axillary lymph nodes is of crucial importance for the staging of breast cancer. As a consequence, an accurate segmentation of the nodes reached by cancer can constitute a precious help for computer-aided diagnosis. However, due to the size of axillary lymph nodes in PET/CT images, and to the low resolution of PET data where their abnomal metabolic hyperactivity may be observed, segmentation remains a challenging task. We investigate the relevance of considering axillary lymph nodes segmentation from PET/CT images, based on Convolutional Neural Networks (CNNs). To this end, our initial working hypotheses were twofold: first, taking advantage of both anatomical information from CT, for detecting the nodes, and from functional information from PET for detecting the inflammatory ones; second, considering region-based attributes extracted from component-tree analysis of PET images in order to enrich the information natively carried by PET, with features that can hardly be inferred by CNNs directly from the images. We describe our first results, and discuss about the validity of these working hypotheses.