In recent years, with the continuous development of deep learning, researchers have proposed many fully supervised deep learning models for diffuse large B-cell lymphoma (DLBCL) segmentation in whole-body PET/CT images. However, as DLBCL belongs to systemic multiple lymphomas and the tumor sizes as well as numbers vary greatly between patients, accurate DLBCL annotation needed by fully supervised deep learning is time-consuming and requires a lot of manual marking work. In this paper, a one-shot learning method is proposed for DLBCL segmentation in whole-body PET/CT images. The proposed method exploits a single lesion randomly selected from each patient' PET/CT images to adapt to the patient’s pathological condition and support lymphoma detection in this patient. Global and local features of the support lesion are fused with features of query images to facilitate lymphoma detection in the two-stage Mask R-CNN framework. Experimental results show that when there are insufficient amounts of training data, the proposed method performs better than the Mask R-CNN.