Fang Dong, Huang Yigeng, Li Suwen, Shi Chen, Bao Junjun, Du Dandan, Xuan Lanlan, Ye Leping, Zhang Yanping, Zhu ChengLin, Zheng Hailun, Shi Zhenwang, Mei Qiao, Wang Huanqin
Department of Gastroenterology, The First Affiliated Hospital of Anhui Medical University, Hefei, 230022, Anhui Province, China.
Department of Gastroenterology, The Second People's Hospital of Hefei, Hefei, 230011, Anhui Province, China.
BMC Cancer. 2025 Mar 18;25(1):495. doi: 10.1186/s12885-025-13910-w.
The cytological diagnostic process of EUS-FNA smears is time-consuming and manpower-intensive, and the conclusion could be subjective and controversial. Moreover, the relative lack of cytopathologists has limited the widespread implementation of Rapid on-site evaluation (ROSE) presently. Therefore, this study aimed to establish an AI system for the detection of pancreatic ductal adenocarcinoma (PDAC) based on EUS-FNA cytological images.
We collected 3213 unified magnification images of pancreatic cell clusters from 210 pancreatic mass patients who underwent EUS-FNA in four hospitals. A semi-supervised CNN (SSCNN) system was developed to distinguish PDAC from Non-PDAC. The internal and external verifications were adopted and the diagnostic accuracy was compared between different seniorities of cytopathologists. 33 images of "Atypical" diagnosed by expert cytopathologists were selected to analyze the consistency between the system and definitive diagnosis.
The segmentation indicators Mean Intersection over Union (mIou), precision, recall and F1-score of SSCNN in internal and external testing sets were 88.3%, 93.21%,94.24%, 93.68% and 87.75%, 93.81%, 93.14%, 93.48% successively. The PDAC classification indicators of the SSCNN model including area under the ROC curve (AUC), accuracy, sensitivity, specificity, PPV and NPV in the internal testing set were 0.97%, 0.95%, 0.94%, 0.97%, 0.98%, 0.91% respectively, and 0.99%, 0.94%, 0.94%, 0.95%, 0.99%, 0.75% correspondingly in the external testing set. The diagnostic accuracy of senior, intermediate and junior cytopathologists was 95.00%, 88.33% and 76.67% under the binary diagnostic criteria of PDAC and non-PDAC. In comparison, the accuracy of the SSCNN system was 90.00% in the dataset of man-machine competition. The accuracy of the SSCNN model was highly consistent with senior cytopathologists (Kappa = 0.853, P = 0.001). The accuracy, sensitivity and specificity of the system in the classification of "atypical" cases were 78.79%, 84.20% and 71.43% respectively.
Not merely tremendous preparatory work was drastically reduced, the semi-supervised CNN model could effectively identify PDAC cell clusters in EUS-FNA cytological smears which achieved analogically diagnostic capability compared with senior cytopathologists, and showed outstanding performance in assisting to categorize "atypical" cases where manual diagnosis is controversial.
This study was registered on clinicaltrials.gov, and its unique Protocol ID was PJ-2018-12-17.
超声内镜引导下细针穿刺（EUS-FNA）涂片的细胞学诊断过程耗时且人力密集，结论可能主观且存在争议。此外，目前细胞病理学家相对短缺限制了快速现场评估（ROSE）的广泛应用。因此，本研究旨在基于EUS-FNA细胞学图像建立一种用于检测胰腺导管腺癌（PDAC）的人工智能系统。
我们收集了来自四家医院接受EUS-FNA的210例胰腺肿块患者的3213张胰腺细胞团统一放大图像。开发了一种半监督卷积神经网络（SSCNN）系统以区分PDAC和非PDAC。采用内部和外部验证，并比较不同资历细胞病理学家的诊断准确性。选择专家细胞病理学家诊断为“非典型”的33张图像分析系统与最终诊断之间的一致性。
SSCNN在内部和外部测试集中的分割指标平均交并比（mIou）、精确率、召回率和F1分数分别为88.3%、93.21%、94.24%、93.68%和87.75%、93.81%、93.14%、93.48%。SSCNN模型在内部测试集中的PDAC分类指标包括ROC曲线下面积（AUC）、准确率、敏感性、特异性、阳性预测值和阴性预测值分别为0.97%、0.95%、0.94%、0.97%、0.98%、0.91%，在外部测试集中相应为0.99%、0.94%、0.94%、0.95%、0.99%、0.75%。在PDAC和非PDAC的二元诊断标准下，高级、中级和初级细胞病理学家的诊断准确率分别为95.00%、88.33%和76.67%。相比之下，在人机竞赛数据集中SSCNN系统的准确率为90.00%。SSCNN模型的准确率与高级细胞病理学家高度一致（Kappa = 0.853，P = 0.001）。该系统在“非典型”病例分类中的准确率、敏感性和特异性分别为78.79%、84.20%和71.43%。
半监督卷积神经网络模型不仅大幅减少了大量的前期准备工作，还能有效识别EUS-FNA细胞学涂片中的PDAC细胞团，与高级细胞病理学家相比具有类似的诊断能力，并且在辅助分类存在人工诊断争议的“非典型”病例方面表现出色。
本研究在clinicaltrials.gov上注册，其唯一的协议编号为PJ-2018-12-17。
