Methods: The proprietary DL-CAD system (MIPNOD 1.0) was pre-trained on the public LIDC-IDRI dataset with 888 CT scans. There were 360 low-dose CT scans that were retrospectively collected from a Chinese lung cancer screening project and evaluated independently by radiologists in a double reading fashion and the DL-CAD system. An extra senior radiologist checked all the results and made the consensus as to the reference standard. Free-response Receiver operating characteristic analysis was applied to assess the detection performance of the DL-CAD system.
Results: After making the consensus, there were 262 nodules in 196 participants and 164 scans without nodules. The DL-CAD system achieved a sensitivity of 89.3% with one false positive per scan, while radiologists had a sensitivity of 76.0% for detection during double reading. Among all undetected nodules, only two were missed by both radiologists and the DL-CAD system. The comparison of detection performance between the DL-CAD system and radiologists in nodule types was as follows. (1) solid nodules (89.7% vs 76.8%; P = 0.003). (2) part-solid nodules (90.9% vs 77.3%; P = 0.375). (3) non-solid nodules (87.3% vs 72.7%; P = 0.134).
Conclusion: The DL-CAD system showed good performance of pulmonary nodule detection in a Chinese population and could potentially provide assistance for radiologists in lung cancer screening programs.
|Journal||European respiratory journal|
|Publication status||Published - 28 Oct 2020|
|Event||ERS International Congress 2020 - Virtual|
Duration: 7 Sep 2020 → 9 Sep 2020
- deep learning
- Lung cancer
- screening program
- lung nodule detection