Feng Yadong, Liang Yan, Li Peng, Long Qigang, Song Jie, Li Mengjie, Wang Xiaofen, Cheng Cui-E, Zhao Kai, Ma Jifeng, Zhao Lingxiao
Department of Gastroenterology, Zhongda Hospital Southeast University, 87 Dingjiaqiao Street, Nanjing, 210009, China.
Department of Gastroenterology, the Affiliated Changshu Hospital of Nantong University, Changshu No. 2 People's Hospital, 18 Taishan Road, Suzhou, 215500, China.
Discov Oncol. 2023 May 19;14(1):73. doi: 10.1007/s12672-023-00694-3.
The use of artificial intelligence (AI) assisted white light imaging (WLI) detection systems for superficial esophageal squamous cell carcinoma (SESCC) is limited by training with images from one specific endoscopy platform.
In this study, we developed an AI system with a convolutional neural network (CNN) model using WLI images from Olympus and Fujifilm endoscopy platforms. The training dataset consisted of 5892 WLI images from 1283 patients, and the validation dataset included 4529 images from 1224 patients. We assessed the diagnostic performance of the AI system and compared it with that of endoscopists. We analyzed the system's ability to identify cancerous imaging characteristics and investigated the efficacy of the AI system as an assistant in diagnosis.
In the internal validation set, the AI system's per-image analysis had a sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) of 96.64%, 95.35%, 91.75%, 90.91%, and 98.33%, respectively. In patient-based analysis, these values were 90.17%, 94.34%, 88.38%, 89.50%, and 94.72%, respectively. The diagnostic results in the external validation set were also favorable. The CNN model's diagnostic performance in recognizing cancerous imaging characteristics was comparable to that of expert endoscopists and significantly higher than that of mid-level and junior endoscopists. This model was competent in localizing SESCC lesions. Manual diagnostic performances were significantly improved with the assistance by AI system, especially in terms of accuracy (75.12% vs. 84.95%, p = 0.008), specificity (63.29% vs. 76.59%, p = 0.017) and PPV (64.95% vs. 75.23%, p = 0.006).
The results of this study demonstrate that the developed AI system is highly effective in automatically recognizing SESCC, displaying impressive diagnostic performance, and exhibiting strong generalizability. Furthermore, when used as an assistant in the diagnosis process, the system improved manual diagnostic performance.
人工智能（AI）辅助白光成像（WLI）检测系统用于浅表性食管鳞状细胞癌（SESCC）的应用受到来自一个特定内镜平台图像训练的限制。
在本研究中，我们使用来自奥林巴斯和富士胶片内镜平台的WLI图像开发了一个带有卷积神经网络（CNN）模型的AI系统。训练数据集由来自1283例患者的5892张WLI图像组成，验证数据集包含来自1224例患者的4529张图像。我们评估了AI系统的诊断性能，并将其与内镜医师的诊断性能进行比较。我们分析了该系统识别癌性影像特征的能力，并研究了AI系统作为诊断辅助工具的有效性。
在内部验证组中，AI系统的逐图像分析的灵敏度、特异度、准确度、阳性预测值（PPV）和阴性预测值（NPV）分别为96.64%、95.35%、91.75%、90.91%和98.33%。在基于患者的分析中，这些值分别为90.17%、94.34%、88.38%、89.50%和94.72%。外部验证组的诊断结果也很理想。CNN模型在识别癌性影像特征方面的诊断性能与专家内镜医师相当，且显著高于中级和初级内镜医师。该模型能够胜任SESCC病变的定位。在AI系统的辅助下，人工诊断性能有显著提高，尤其是在准确度（75.12%对84.95%，p = 0.008）、特异度（63.29%对76.59%，p = 0.017）和PPV（64.95%对75.23%，p = 0.006）方面。
本研究结果表明，所开发的AI系统在自动识别SESCC方面非常有效，显示出令人印象深刻的诊断性能，并具有很强的通用性。此外，当该系统用作诊断过程中的辅助工具时，提高了人工诊断性能。
