A cohort of in-house 195 patients with breast cancer who underwent PET-CT, between 2010 and 2020 was selected. All suspicious lymph nodes in all cases were annotated by a radiologist based on the corresponding report. First, a pre-trained lesion segmentation model was developed using the publicly available AutoPET dataset. Based on this pre-trained model, separate segmentation models were trained to segment suspicious positive axillary lymph nodes and parasternal lymph nodes. For axillary lymph nodes, the region of interest was cropped to 64x64x64 and fed into a convolutional neural network to predict the stage according to the number of positive lymph nodes. Finally, the results of the parasternal lymph node segmentation model were combined to give a more refined staging (Figure 1). The Dice coefficient, area under the curve (AUC) and accuracy were used to evaluate the performance of the model in segmentation and classification tasks, respectively.