Artificial intelligence deep learning software for segmentation of mediastinal and hilar lymph nodes on plain CT: a retrospective study in a cancer population

Abstract Background Accurate identification and characterization of lymph nodes (LNs) are essential in cancer staging and treatment planning. While artificial intelligence (AI) has shown potential in detecting lymphadenopathy on contrast-enhanced CT scans, its performance on non-contrast (plain) CT scans remains less explored. This retrospective study aimed to assess the efficacy of AI for segmenting LNs on plain CT in patients with lung or gastrointestinal cancer. Eligible cases were selected based on plain CT scans with 5 mm slice thickness, showing mediastinal or hilar LNs with a short-axis diameter of ≥ 10 mm. A total of 144 patients (105 men, 39 women; mean age, 76 years) were included, comprising 342 LNs. A commercially available deep learning AI software was used for segmentation, with the ground-truth defined by two radiologists’ consensus on LNs measuring ≥ 10 mm. Results The AI software showed a sensitivity of 83.9%, with no statistically significant difference between mediastinal and hilar LNs ( p = 0.84). The false-positive rate was 0.465 per patient. The Dice coefficient was 0.908, and there was no statistically significant difference between the mediastinum and hilum ( p = 0.19). The sensitivity decreased in larger LNs, showing an inverse correlation ( r = −0.122, p < 0.01). Conclusions This study demonstrated the robustness of the AI software in the segmentation of LNs in the mediastinum and hilum on plain CT scans.