Division of Gastroenterology and Hepatology, Department of Internal Medicine, Korea University College of Medicine, 73 Goryeodae-ro, Seongbuk-gu, Seoul, 02841, Republic of Korea.
Center for Bionics, Korea Institute of Science and Technology (KIST), 5, Hwarang-ro 14-gil, Seongbuk-gu, Seoul, 02792, Republic of Korea.
Sci Rep. 2021 Mar 5;11(1):5311. doi: 10.1038/s41598-021-84299-2.
The treatment plan of colorectal neoplasm differs based on histology. Although new endoscopic imaging systems have been developed, there are clear diagnostic thresholds and requirements in using them. To overcome these limitations, we trained convolutional neural networks (CNNs) with endoscopic images and developed a computer-aided diagnostic (CAD) system which predicts the pathologic histology of colorectal adenoma. We retrospectively collected colonoscopic images from two tertiary hospitals and labeled 3400 images into one of 4 classes according to the final histology: normal, low-grade dysplasia, high-grade dysplasia, and adenocarcinoma. We implemented a CAD system based on ensemble learning with three CNN models which transfer the knowledge learned from common digital photography images to the colonoscopic image domain. The deep learning models were trained to classify the colorectal adenoma into these 4 classes. We compared the outcomes of the CNN models to those of two endoscopist groups having different years of experience, and visualized the model predictions using Class Activation Mapping. In our multi-center study, our CNN-CAD system identified the histology of colorectal adenoma with as sensitivity 77.25%, specificity of 92.42%, positive predictive value of 77.16%, negative predictive value of 92.58% averaged over the 4 classes, and mean diagnostic time of 0.12 s per image. Our experiments demonstrate that the CNN-CAD showed a similar performance to that of endoscopic experts and outperformed that of trainees. The model visualization results also showed reasonable regions of interest to explain the classification decisions of CAD systems. We suggest that CNN-CAD system can predict the histology of colorectal adenoma.
结直肠肿瘤的治疗方案因组织学而异。尽管已经开发出了新的内镜成像系统，但在使用它们时仍存在明确的诊断阈值和要求。为了克服这些局限性，我们使用内镜图像对卷积神经网络（CNN）进行了训练，并开发了一种计算机辅助诊断（CAD）系统，该系统可预测结直肠腺瘤的病理组织学。我们回顾性地从两家三级医院收集了结肠镜图像，并根据最终组织学将 3400 张图像标记为以下 4 个类别之一：正常、低级别异型增生、高级别异型增生和腺癌。我们基于集成学习实现了一个 CAD 系统，该系统使用三个 CNN 模型，将从普通数字摄影图像中学到的知识转移到结肠镜图像领域。深度学习模型经过训练可将结直肠腺瘤分为这 4 个类别。我们将 CNN 模型的结果与具有不同经验的两组内镜医生的结果进行了比较，并使用 Class Activation Mapping 可视化了模型预测。在我们的多中心研究中，我们的 CNN-CAD 系统以平均 4 个类别的敏感性 77.25%、特异性 92.42%、阳性预测值 77.16%、阴性预测值 92.58%来识别结直肠腺瘤的组织学，平均每张图像的诊断时间为 0.12 秒。我们的实验表明，CNN-CAD 与内镜专家的表现相当，优于培训师。模型可视化结果还显示了合理的感兴趣区域，以解释 CAD 系统的分类决策。我们建议 CNN-CAD 系统可以预测结直肠腺瘤的组织学。
