The evaluation of mediastinal nodal involvement in oncologic patients is of the utmost importance, being the lymph nodes a possible site of metastatic seeding in lung cancer and other thoracic and extra-thoracic tumors. Nodal assessment is generally performed on CT scans by radiologists as part of the radiologic workup through the analysis of medical images, employing different diagnostic criteria with Response Evaluation Criteria In Solid Tumors (RECIST) accounting for the most used. Recent studies suggested that the radiomic/quantitative analysis of lymph nodes may outperform the standard evaluation [1]. Even in the standard RECIST evaluation, it has been demonstrated that there is an important variability between measurements performed by different radiologists and in some cases also the lesions that are considered as target lesions may differ. Conversely, automated systems have been found to be more reproducible, allowing higher consistency [2]. Our aim was to train a neural network for the recognition and automatic segmentation of mediastinal lymph nodes, evaluating the impact of improved manual segmentation and fine-tuning of the models’ hyperparameters on overall performance.