人工智能技术对肺部亚实性结节检出及定性 诊断效能评估

Evaluation of Artificial Intelligence Technology in the Detection and Qualitative Diagnosis of Pulmonary Subsolid Nodules

摘要: 目的：评估人工智能(Artificial intelligence, AI)在肺部亚实性结节(Sub-solid nodule, SSN) CT影像诊断中的效能，分析AI对不同病理亚型的鉴别能力及误诊漏诊特征，为临床辅助诊断提供更精准的参考依据。方法：回顾性分析2019年1月~2025年5月于天津市北辰医院接诊的280例SSN患者资料，以手术病理结果为金标准，明确AI单独阅片为算法独立输出诊断结果，设置人工单独阅片、AI单独阅片、医生 + AI联合阅片三组，比较三组对纯磨玻璃结节(Pure ground glass nodule, pGGN)、部分实性结节(Partial solid nodules, PSN)及不同病理亚型的良恶性检出结果，统计分析三组诊断效能；通过Kappa检验评估三组与金标准的一致性，定性分析AI及人工阅片的误诊、漏诊病例特征，并结合典型误诊图像佐证结论。结果：AI单独阅片对pGGN、PSN的CT图像判读的准确度、特异度、敏感度均高于人工单独阅片，医生 + AI联合阅片诊断效能为三组最优；AI单独阅片对pGGN、PSN良恶性检出结果与手术病理金标准一致性较好(Kappa值 = 0.846、0.828，P < 0.01)，人工单独阅片一致性一般(Kappa值 = 0.412、0.400，P < 0.01)，医生 + AI联合阅片一致性最优(Kappa值 = 0.915、0.902，P < 0.01)。AI对不典型腺瘤样增生(Atypical adenomatous hyperplasia, AAH)的过度诊断及微浸润腺癌(Microinvasive adenocarcinoma, MIA)与浸润性腺癌(Invasive adenocarcinoma, IAC)的鉴别偏差为主要误诊类型，漏诊病例约5%，主要因病灶过小、位置特殊及密度过低导致；人工阅片漏诊误诊率高主要源于主观经验差异、视觉疲劳及细微特征识别不足等。结论：相较于传统的人工阅片方式，AI辅助技术展现出更优的诊断效能，医生 + AI联合阅片更符合临床实际应用场景，可显著提升诊断准确性。AI在SSN病理亚型鉴别中存在一定局限性，其误诊漏诊具有明确的特征性，临床应用中需结合人工阅片进行综合判断，该技术可为放射科医师进行SSN性质判断提供重要辅助价值。

Abstract: Objective: To evaluate the effectiveness of artificial intelligence (AI) in CT image diagnosis of sub-solid nodules (SSN) of the lung, and analyze the ability of AI to distinguish different pathological subtypes and the characteristics of misdiagnosis and missed diagnosis, so as to provide a more accurate reference for clinical auxiliary diagnosis. Methods: The data of 280 patients with SSN who were treated in Tianjin Beichen Hospital from January 2019 to May 2025 were retrospectively analyzed. Taking the surgical and pathological results as the gold standard, it was clear that AI individual film reading was the algorithm to independently output the diagnostic results. Three groups were set up, including manual individual film reading, AI individual film reading, and doctor + AI joint film reading. The effects of the three groups on pure ground glass nodules (pGGN), partial solid nodules were compared (PSN) and the benign and malignant detection results of different pathological subtypes, and the diagnostic efficacy of the three groups was statistically analyzed. The consistency between the three groups and the gold standard was evaluated by the Kappa test. The characteristics of misdiagnosed and missed cases of AI and manual film reading were qualitatively analyzed, and the conclusion was supported by typical misdiagnosed images. Results: The accuracy, specificity and sensitivity of AI alone in the interpretation of CT images of pGGN and PSN were higher than those of manual reading alone. The diagnostic efficiency of doctor + AI combined reading was the best in the three groups (P < 0.01). The consistency of manual individual film reading was general (Kappa value = 0.412, 0.400, P < 0.01). The consistency of doctor + AI joint film reading was the best (Kappa = 0.915, 0.902, P < 0.01). AI overdiagnosis of atypical adenomatous hyperplasia (Aah) and microinvasive adenocarcinoma (MIA) and invasive adenocarcinoma, IAC) was the main misdiagnosis type, and the missed cases were about 5%, which was mainly caused by the small lesion, special location and low density; The high rate of missed diagnosis and misdiagnosis in manual film reading is mainly due to the difference of subjective experience, visual fatigue and insufficient recognition of subtle features. Conclusion: Compared with the traditional manual film reading method, AI-assisted technology shows better diagnostic efficiency, and the doctor + AI joint film reading is more in line with the clinical application scenario, which can significantly improve the diagnostic accuracy. AI has certain limitations in the identification of SSN pathological subtypes. Its misdiagnosis and missed diagnosis have clear characteristics. In clinical application, it needs to be combined with manual film reading for comprehensive judgment. This technology can provide important auxiliary value for radiologists to judge the nature of SSN.