Tung Chun-Liang, Chang Han-Cheng, Yang Bo-Zhi, Hou Keng-Jen, Tsai Hung-Hsu, Tsai Cheng-Yu, Yu Pao-Ta
Department of Pathology, Ditmanson Medical Foundation Chiayi Christian Hospital, Chiayi, Taiwan; Department of Health and Nutrition Biotechnology, Asia University, Taichung, Taiwan.
Department of Computer Science & Information Engineering, National Chung Cheng University, Chiayi, Taiwan; Information Technology Department, Ditmanson Medical Foundation Chiayi Christian Hospital, Chiayi, Taiwan.
J Formos Med Assoc. 2022 Dec;121(12):2457-2464. doi: 10.1016/j.jfma.2022.05.004. Epub 2022 Jun 3.
The accuracy of histopathology diagnosis largely depends on the pathologist's experience. It usually takes over 10 years to cultivate a senior pathologist, and small numbers of them lead to a high workload for those available. Meanwhile, inconsistent diagnostic results may arise among different pathologists, especially in complex cases, because diagnosis based on morphology is subjective. Computerized analysis based on deep learning has shown potential benefits as a diagnostic strategy.
This research aims to automatically determine the location of gastric cancer (GC) in the images of GC slices through artificial intelligence. We use image data from a regional teaching hospital in Taiwan for training. We collect images of patients diagnosed with GC from January 1, 2019 to December 31, 2020. In this study, scanned images are used to dissect 13,600 images from 50 different patients with GC sections whose GC sections are stained with hematoxylin and eosin (H&E stained) through a whole slide scanner, the scanned images from 50 different GC slice patients are divided into 80% training combinations, 2200 images of 40 patients are trained. The remaining 20%, totaling 10 people, are validated from a test set of 550 images.
The validation results show that 91% of the correct rates are interpreted as GC images through deep learning. The sensitivity, specificity, PPV, and NPV were 84.9%, 94%, 87.7%, and 92.5%, respectively. After creating a 3D model through the grayscale value, the position of the GC is completely marked by the 3D model. The purpose of this research is to use artificial intelligence (AI) to determine the location of the GC in the image of GC slice.
In patients undergoing pancreatectomy for pancreatic cancer, intraoperative infusion of lidocaine did not improve overall or disease-free survival. Reduced formation of circulating NETs was absent in pancreatic tumour tissue.
For AI to assist pathologists in daily practice, to help a pathologist making a definite diagnosis is not the prime purpose at present time. The benefits could come from cancer screening and double-check quality control in a heavy workload which could distract the attention of pathologist during the time constraint examination process. We propose a two-steps method to identify cancerous areas in endoscopic gastric biopsy slices via deep learning. Then a 3D model is used to further mark all the positions of GC in the picture, and the model overcomes the problem that deep learning can't catch all GC.
组织病理学诊断的准确性很大程度上取决于病理学家的经验。培养一名资深病理学家通常需要10年以上时间，而人数较少导致现有病理学家工作量很大。同时，不同病理学家之间可能会出现诊断结果不一致的情况，尤其是在复杂病例中，因为基于形态学的诊断是主观的。基于深度学习的计算机分析已显示出作为一种诊断策略的潜在益处。
本研究旨在通过人工智能自动确定胃癌（GC）切片图像中胃癌的位置。我们使用台湾一家地区教学医院的图像数据进行训练。我们收集了2019年1月1日至2020年12月31日期间被诊断为胃癌的患者的图像。在本研究中，通过全玻片扫描仪使用扫描图像从50名不同的胃癌切片患者中剖析出13600张图像，这些患者的胃癌切片用苏木精和伊红染色（H&E染色），将来自50名不同胃癌切片患者的扫描图像分为80%的训练组合，对40名患者的2200张图像进行训练。其余20%，共10人，从550张图像的测试集中进行验证。
验证结果表明，通过深度学习将91%的正确率解释为胃癌图像。敏感性、特异性、阳性预测值和阴性预测值分别为84.9%、94%、87.7%和92.5%。通过灰度值创建3D模型后，胃癌的位置由3D模型完全标记。本研究的目的是使用人工智能（AI）确定胃癌切片图像中胃癌的位置。
在接受胰腺癌胰腺切除术的患者中，术中输注利多卡因并未改善总体生存率或无病生存率。胰腺肿瘤组织中循环中性粒细胞胞外陷阱的形成减少。
目前，人工智能辅助病理学家日常工作，帮助病理学家做出明确诊断并非首要目的。其益处可能来自癌症筛查以及在繁重工作量下的双重检查质量控制，这在时间紧迫的检查过程中可能会分散病理学家的注意力。我们提出一种两步法，通过深度学习识别内镜胃活检切片中的癌区。然后使用3D模型进一步标记图片中胃癌的所有位置，该模型克服了深度学习无法捕捉所有胃癌的问题。
