Vision Transformer-based models, such as Swin-Unet, require large volumes of training data. Self-supervised learning (SSL) mitigates this problem by pretraining models on unlabeled data. This paper presents a novel SSL method, Axial Distance Prediction (ADP), that exploits the volumetric nature of three-dimensional medical image data by predicting the distance between pairs of axial slices in CT scans based on their embeddings. We demonstrated this approach by pretraining a Swin-Unet model for lung nodule segmentation on the LIDC-IDRI data set and compared its performance against state-of-the-art SSL methods: DCL, BYOL and SimMIM. Our experiments showed that ADP is a competitive SSL method for lung nodule segmentation, especially for very small training sets. Moreover, we showed that our method is less computationally expensive than state-of-the-art SSL approaches (DCL and BYOL).