基于深度学习的人工智能辅助肺腺癌胸水脱落细胞学诊断的方法

目的 应用深度学习模型对胸水脱落肺腺癌细胞进行检测和分类，探讨人工智能辅助肺癌细胞病理诊断的可行性。方法 收集2019年3月至2021年12月南通大学附属医院、上海交通大学附属胸科医院和复旦大学附属中山医院的肺腺癌胸水标本110例，非癌性胸水标本20例为对照组。采用常规法和单细胞分离液处理技术2种方法分离细胞标本，并进行液基制片和苏木精-伊红（H-E）染色，全切片数字扫描使细胞图像数字化后保存为数字文件，由人工智能辅助诊断。经过裁切与图像预处理后，使用LabelImg软件对胸水细胞进行标记，打方框并标注细胞类型，选用较典型细胞样本，分别标记淋巴细胞、间皮细胞和腺癌细胞，共标记800张图像用于训练。采用Yolo V4模型对疑似与确诊肺癌细胞进行训练，采用Inception V3模型对不同分类细胞进行训练，取另外250张图像进行测试。结果 训练后的Yolo V4模型能够对胸水脱落细胞H-E染色涂片中疑似+确诊肺腺癌细胞进行识别标注，全类平均正确率（mAP）为20%；训练后的Inception V3模型对胸水脱落细胞病理图像中单个细胞分割后的淋巴细胞、间皮细胞、疑似+确诊肺腺癌细胞进行分类，...

Deep learning for artificial intelligence aided cytological diagnosis on exfoliated adenocarcinoma cells of lung in pleural effusion

Objective To explore the feasibility of artificial intelligence (AI)-assisted pathological diagnosis on lung adenocarcinoma cells by using deep learning model. Methods From March 2019 to December 2021, 110 specimens of lung adenocarcinoma were collected from the Affiliated Hospital of Nantong University, the Chest Hospital of Shanghai Jiao Tong University and Zhongshan Hospital, Fudan University and 20 specimens of non-cancerous pleural effusion as the control group. Two methods of cell separation (routine technology and new treatment technology with single cell separation solution) were used. Then making liquid based thin layer cell slides, H-E staining, digital scanning of whole slide image (WSI) and saving as digital files. Afterwards AI-assisted diagnosis was performed. After cutting and image pretreatment, LabelImg software was used to label pleural fluid cells, box and note cell types. Typical cell samples were selected to label lymphocytes, mesothelial cells and adenocarcinoma cells respectively. A total of 800 images were labeled for training. And then machine learning, suspected and confirmed lung adenocarcinoma cells were trained with Yolo V4 model, cells of different classifications were trained with Inception V3 model, and another 250 images were taken for testing. Results The trained Yolo V4 model could identify suspected + confirmed lung cancer cells in HE staining smears of pleural fluid cells (mAP 20%). The trained Inception V3 model can classify lymphatic, mesothelial, and suspected + confirmed lung adenocarcinoma cells segmented by a single cell in the pathological images from exfoliated cells of pleural fluid with an accuracy of 98%. Single cell separation solution can increase the number of cancer cells that could be clearly labeled, and increase the efficiency and accuracy of single cell target recognition. Conclusion Based on deep learning method, the AI model can detect and classify lung adenocarcinoma cells in exfoliated pleural fluid cells, and can be used to assist the pathologic diagnosis of lung cancer. Improving the efficiency of cell separation and unified standardized preparation are helpful to promote clinical application.