AI辅助软件对于提高不同年资规培医师肺结节检测能力的影响

Improving the identification of pulmonary nodules:An AI-assisted training for physicians with varied clinical experiences

目的评估人工智能(AI)辅助软件对低、高不同年资规培医师在提高胸部CT肺结节的检测效能中的应用价值。方法回顾性收集行胸部CT检查有1枚及以上肺结节且结节直径为1 mm^3 cm的181例病人。分别在低、高年资规培医师独立和借助AI软件辅助检测的4种情况下检测不同密度(实性、纯磨玻璃、混合磨玻璃和钙化)及不同大小(<5 mm、5~10 mm和>10 mm)肺结节的敏感度和假阳性结节数。分别采用卡方检验和配对t检验或Wilcoxon秩和检验比较低、高年资规培医师检测结节的敏感度和假阳性结节数。结果对于实性结节及<5 mm结节的检测,高年资医师的敏感度高于低年资医师(均P<0.05);在AI辅助下阅片时,高年资医师检测实性结节、纯磨玻璃结节(pGGN)、钙化结节及<5 mm结节、5~10 mm结节的敏感度均高于低年资医师(均P<0.05)。无论是独立阅片还是AI辅助阅片,低、高年资医师对>10 mm肺结节假阳性结节数的检测差异均无统计学意义(P>0.05),而对于检测其余肺结节的假阳性结节数,低年资医师均多于高年资医师(均P<0.05)。对于钙化结节、>10 mm肺结节的检测,低年资医师在独立阅片和AI辅助阅片下的敏感度差异无统计学意义(P>0.05),而对于其余肺结节的检测,AI辅助阅片的敏感度均较独立阅片时有明显提高(均P<0.05);并且低年资医师在有无AI辅助的阅片对检测>10 mm肺结节假阳性结节数的差异无统计学意义(P>0.05),而在AI辅助下,其检测实性结节、钙化结节、<5 mm结节及5~10 mm结节的假阳性结节数均较独立阅片减少,但检测pGGN和混合磨玻璃结节(mGGN)的假阳性结节数增多(均P<0.05)。高年资医师有无AI辅助阅片,对>10 mm肺结节的检测敏感度及假阳性结节数的差异均无统计学意义(均P>0.05),而在AI辅助阅片下,其检测其余肺结节的敏感度均较独立阅片时有显著提高,且假阳性结节数均不同程度减少(均P<0.05)。结论 AI辅助软件能够有效提高规培医师对CT上不同肺结节的检测效能,降低漏诊风险,但低年资规培医师检测的假阳性结节数仍较多。

Objective To evaluate the application value of artificial intelligence(AI) assisted software in improving the detection efficiency of pulmonary nodules on CT by physicians with varied clinical experiences. Methods A total of 181 patients with one or more pulmonary nodules were retrospectively collection. The diameters of nodules were 1 mm to 3 cm.All the patients underwent chest CT examination. Without and with assistance of AI software, the sensitivity in identifying pulmonary nodules with different densities(solid, pure ground glass, mixed ground glass, and calcification) and different sizes(<5 mm, 5-10 mm, and >10 mm) and the number of false positive nodules by residents and interns were calculated. The Chisquare test and paired t test or Wilcoxon rank sum test were used to compare the sensitivities and the numbers of false positive nodules between before and after AI assistance. Results For the solid nodules and <5 mm nodules, the sensitivity in identifying the nodules by residents was higher than that by interns(all P<0.05). When under the assistance of AI, the sensitivity in identifying the solid nodules, pure ground glass nodules, calcified nodules, <5 mm nodules, and 5-10 mm nodules by residents were higher than that by interns(all P<0.05). With or without AI assistance, the number of false positive nodules of >10 mm nodules was not significantly different in both interns and residents(all P>0.05), but for identifying the other nodules, the number of false positive nodules in interns was higher than that in residents(all P <0.05). For the calcification nodules and >10 mm pulmonary nodules, AI assistance has no significant effect on the sensitivity in identifying the nodules by interns(P>0.05), but for other nodules, the sensitivity with AI assistance were significantly higher than that of independent reading(all P<0.05). AI assistance had no significant effect on the number of false positive nodules of >10 mm nodules in interns(P>0.05), and the number of false positive nodules in identifying the solid nodules, calcification nodules, <5 mm nodules, and 5-10 mm nodules were reduced by interns with AI assistance, however, the number of false positive nodules in identifying the pure ground glass nodules and mixed ground glass were increased(all P<0.05). In residents, AI assistance had no significantly effect on the sensitivity and the number of false positive nodules in identifying >10 mm nodules(all P>0.05), but when resident reading under AI-assistance, the sensitivities in identifying the other nodules were all significantly improved and the numbers of false positive nodules were reduced(all P <0.05). Conclusion AI-assisted software can greatly improve the efficiency of intern and resident in detecting different pulmonary nodules on CT, thereby reducing the risk of missed diagnosis, but for interns, the number of false positive nodules is still high.