肺癌人工智能细胞病理诊断系统的研发及诊断价值探讨

Development and diagnostic value of artificial intelligence cytopathological diagnosis system for lung cancer

背景传统细胞病理诊断肺癌虽具有优势,但受医师主观经验和工作负荷影响较大。以深度学习算法模型为代表的新一代人工智能,能够自动提取和归纳医学图像中的特征,在智能诊断中展现出显著优势。目的结合人工智能和数字病理学新技术,研发适用于肺癌的人工智能细胞病理诊断系统并对其诊断价值进行评价。方法选取解放军总医院第一医学中心2021年5月—2023年7月临床拟诊肺癌患者533例,其中最终病理确诊肺癌354例(包括腺癌98例、鳞癌140例、小细胞癌116例),非肺癌179例。将选取病例的气管镜活检标本及胸腔积液标本进行涂片、染色、扫描。使用随机选取的340例样本(肺癌病例229例和非肺癌病例111例)的数字病理切片分别对备选的检测模型和分类模型进行训练、验证及测试,根据测试结果,择优选取YOLO v7检测模型及Vision Transformer分类模型为基本架构初步建立肺癌人工智能细胞病理诊断系统。利用训练过的人工智能细胞病理诊断系统对剩下的193例未经训练的样本进行诊断测试,以病理学诊断结果为标准比较判读结果。结果本研究研发的人工智能细胞病理诊断系统在肺癌诊断中的准确率为91.2%(176/193),敏感度为98.4%(123/125),特异度为77.9%(53/68),阳性预测值为89.1%(123/138),阴性预测值为96.4%(53/55),Youden指数为0.763,Kappa值为0.798。结论人工智能细胞病理诊断系统对肺癌的诊断敏感度和准确率均较高,可有效提高肺癌诊断效率,但该系统仍需进一步优化,从而提高诊断特异度。

Background Although traditional cytopathological diagnosis of lung cancer has its advantages,it is greatly influenced by doctors'subjective experience and workload.The new generation of artificial intelligence,represented by deep learning algorithm models,can automatically extract and summarize features from medical images,demonstrating significant advantages in intelligent diagnosis.Objective To develop an artificial intelligence cytopathological diagnosis system for lung cancer and explore its diagnostic value by combining the artificial intelligence and digital pathology.Methods From May 2021 to July 2023,533 patients with suspected lung cancer were selected from the First Medical Center of Chinese PLA General Hospital.Among them,354 cases were finally diagnosed with lung cancer(including 98 cases of adenocarcinoma,140 cases of squamous carcinoma,and 116 cases of small-cell carcinoma),and another 179 cases were non-lung cancer.The bronchoscopic biopsy specimens and pleural effusion specimens from the selected cases were smeared,stained,and scanned.Using the digital pathological slices from 340 randomly selected samples(including 229 lung cancer cases and 111 non-lung cancer cases),the candidate detection models and classification models were trained,validated,and tested,respectively.Based on the test results,the YOLO v7 detection model and Vision Transformer classification model were selected as the basic structure to initially establish the Artificial Intelligence Cytopathological Diagnosis System for lung cancer.Then the trained Artificial Intelligence Cytopathological Diagnosis System was used to diagnose the remaining 193 untrained samples for validation, and the interpretation results were compared with the pathological diagnosis results as the standard. Results The accuracy of the developed Artificial Intelligence Cytopathological Diagnosis System in lung cancer diagnosis was 91.2% (176/193), with sensitivity of 98.4% (123/125), specificity of 77.9% (53/68), positive predictive value of 89.1% (123/138), and negative predictive value of 96.4% (53/55). The Youden index was 0.763, and Kappa statistic was 0.798. Conclusion The Artificial Intelligence Cytopathological Diagnosis System has high sensitivity and accuracy in the diagnosis of lung cancer, which can effectively improve the efficiency of lung cancer diagnosis. However, the system still needs to be further optimized to enhance its diagnostic specificity.