Department of Gastroenterology, Cancer Institute Hospital, Japanese Foundation for Cancer Research, Tokyo, Japan.
Tada Tomohiro Institute of Gastroenterology and Proctology, Saitama, Japan.
Endoscopy. 2020 Dec;52(12):1077-1083. doi: 10.1055/a-1194-8771. Epub 2020 Jun 8.
We previously reported for the first time the usefulness of artificial intelligence (AI) systems in detecting gastric cancers. However, the "original convolutional neural network (O-CNN)" employed in the previous study had a relatively low positive predictive value (PPV). Therefore, we aimed to develop an advanced AI-based diagnostic system and evaluate its applicability for the classification of gastric cancers and gastric ulcers.
We constructed an "advanced CNN" (A-CNN) by adding a new training dataset (4453 gastric ulcer images from 1172 lesions) to the O-CNN, which had been trained using 13 584 gastric cancer and 373 gastric ulcer images. The diagnostic performance of the A-CNN in terms of classifying gastric cancers and ulcers was retrospectively evaluated using an independent validation dataset (739 images from 100 early gastric cancers and 720 images from 120 gastric ulcers) and compared with that of the O-CNN by estimating the overall classification accuracy.
The sensitivity, specificity, and PPV of the A-CNN in classifying gastric cancer at the lesion level were 99.0 % (95 % confidence interval [CI] 94.6 %-100 %), 93.3 % (95 %CI 87.3 %-97.1 %), and 92.5 % (95 %CI 85.8 %-96.7 %), respectively, and for classifying gastric ulcers were 93.3 % (95 %CI 87.3 %-97.1 %), 99.0 % (95 %CI 94.6 %-100 %), and 99.1 % (95 %CI 95.2 %-100 %), respectively. At the lesion level, the overall accuracies of the O- and A-CNN for classifying gastric cancers and gastric ulcers were 45.9 % (gastric cancers 100 %, gastric ulcers 0.8 %) and 95.9 % (gastric cancers 99.0 %, gastric ulcers 93.3 %), respectively.
The newly developed AI-based diagnostic system can effectively classify gastric cancers and gastric ulcers.
我们之前首次报道了人工智能（AI）系统在检测胃癌方面的有效性。然而，我们之前研究中使用的“原始卷积神经网络（O-CNN）”的阳性预测值（PPV）相对较低。因此，我们旨在开发一种先进的基于 AI 的诊断系统，并评估其在胃癌和胃溃疡分类中的适用性。
我们通过向 O-CNN 添加新的训练数据集（来自 1172 个病变的 4453 个胃溃疡图像）来构建“先进的卷积神经网络（A-CNN）”，该 O-CNN 已经使用 13584 个胃癌和 373 个胃溃疡图像进行了训练。使用独立的验证数据集（来自 100 个早期胃癌的 739 个图像和来自 120 个胃溃疡的 720 个图像），对 A-CNN 在分类胃癌和溃疡方面的诊断性能进行回顾性评估，并通过估计整体分类准确性与 O-CNN 进行比较。
A-CNN 在病变水平上分类胃癌的灵敏度、特异性和 PPV 分别为 99.0%（95%置信区间[CI]94.6%-100%）、93.3%（95%CI87.3%-97.1%）和 92.5%（95%CI85.8%-96.7%），分类胃溃疡的灵敏度、特异性和 PPV 分别为 93.3%（95%CI87.3%-97.1%）、99.0%（95%CI94.6%-100%）和 99.1%（95%CI95.2%-100%）。在病变水平上，O-CNN 和 A-CNN 对胃癌和胃溃疡的总体准确率分别为 45.9%（胃癌 100%，胃溃疡 0.8%）和 95.9%（胃癌 99.0%，胃溃疡 93.3%）。
新开发的基于 AI 的诊断系统可以有效地对胃癌和胃溃疡进行分类。
