Effect of detection algorithm based on deep learning for pulmonary nodules in different lung lobes / 中国医学影像技术

ABSTRACT

Objective: To explore the effect of pulmonary nodules in different lung lobes detection algorithm based on deep learning (DL). Methods: Totally 493 eligible patients with pulmonary nodules on chest CT were included, and pulmonary nodules were labeled. The results of pulmonary nodules detection algorithm based on DL were compared with those of radiologist's labelling, and the match ratios in every lung lobe were counted, respectively. The radiologist finally re-evaluated the nodules that might be detected by algorithm but were missed during the initial inspection. Results: The match ratio of 4.1 -30.0 mm nodules of DL algorithm was 96.05% (73/76), 96.91% (94/97), 96.94% (95/98), 98.59% (70/71), 95.95% (71/74) and 96.30% (26/27) for pulmonary nodules in left upper lobe, left lower lobe, right upper lobe, right middle lobe, right lower lobe and interlobar pleura, respectively (all P>0.05). After re-evaluation, 50.92% (747/1467) of the no matched nodules detected by algorithm were reassigned as true positives. There were statistical differences among the missed nodules on different lobes (all P<0.05). Conclusion: The performance of pulmonary nodule detection algorithm based DL is not affected by nodule locations in terms of pulmonary lobes. The distribution of missed nodules meets the general consensus of medical profession.