Effects of detection algorithm based on deep learning for different size pulmonary nodule / 中国医学影像技术

ABSTRACT

Objective: To explore the diagnostic effects of detection algorithm based on deep learning (DL) on pulmonary nodules with different sizes. Methods: CT images of 344 patients with pulmonary nodules were retrospectively analyzed. The detection rates of the model based on DL for pulmonary nodules with different sizes (relative to the physician's diagnosis) were calculated and compared, and the false positive nodules detected by the model were analyzed. Results: On 344 CT images, physicians diagnosed 710 pulmonary nodules of 0-30 mm. A total of 2 495 candidate pulmonary nodules were detected by the model, among which 675 were true positive relative to the physician's diagnosis. The detection rate of nodules of the model was 95.07% (675/710), of 0-4 mm was 82.80% (77/93), of 0-5 mm was 90.15% (238/264), of 0-6 mm was 92.94% (395/425), of 5-10 mm was 97.94% (381/389), of 10-20 mm was 98.21% (55/56), and of 20-30 mm was 100% (1/1). There was no statistically significant difference of detection rate for pulmonary nodules with different sizes of the model(χ2=21.72, P>0.05). Among the false positive nodules detected by the model, 50.38% (917/1 820) were missed by physicians, and 32.53% (592/1 820) were vascular sections. Conclusion: The overall detection rate of pulmonary nodules of DL model is high (95.07%), which is not affected by the size of nodules.