INTRODUCTION Endoscopic evaluation is crucial for predicting the invasion depth of esophagus squamous cell carcinoma (ESCC) and selecting appropriate treatment strategies. Our study aimed to develop and validate an interpretable artificial intelligence-based invasion depth prediction system (AI-IDPS) for ESCC. METHODS We reviewed the PubMed for eligible studies and collected potential visual feature indices associated with invasion depth. Multicenter data comprising 5,119 narrow-band imaging magnifying endoscopy images from 581 patients with ESCC were collected from 4 hospitals between April 2016 and November 2021. Thirteen models for feature extraction and 1 model for feature fitting were developed for AI-IDPS. The efficiency of AI-IDPS was evaluated on 196 images and 33 consecutively collected videos and compared with a pure deep learning model and performance of endoscopists. A crossover study and a questionnaire survey were conducted to investigate the system's impact on endoscopists' understanding of the AI predictions. RESULTS AI-IDPS demonstrated the sensitivity, specificity, and accuracy of 85.7%, 86.3%, and 86.2% in image validation and 87.5%, 84%, and 84.9% in consecutively collected videos, respectively, for differentiating SM2-3 lesions. The pure deep learning model showed significantly lower sensitivity, specificity, and accuracy (83.7%, 52.1% and 60.0%, respectively). The endoscopists had significantly improved accuracy (from 79.7% to 84.9% on average, P = 0.03) and comparable sensitivity (from 37.5% to 55.4% on average, P = 0.27) and specificity (from 93.1% to 94.3% on average, P = 0.75) after AI-IDPS assistance. DISCUSSION Based on domain knowledge, we developed an interpretable system for predicting ESCC invasion depth. The anthropopathic approach demonstrates the potential to outperform deep learning architecture in practice.

中文翻译：

---

使用放大窄带成像内窥镜预测食管鳞状细胞癌侵袭深度的类人人工智能系统：一项回顾性多中心研究。

引言内镜评估对于预测食管鳞状细胞癌（ESCC）的侵袭深度和选择适当的治疗策略至关重要。我们的研究旨在开发和验证一种可解释的基于人工智能的 ESCC 侵袭深度预测系统 (AI-IDPS)。方法 我们回顾了 PubMed 中符合条件的研究，并收集了与浸润深度相关的潜在视觉特征指数。2016 年 4 月至 2021 年 11 月期间，从 4 家医院收集了 581 名 ESCC 患者的 5,119 张窄带成像放大内窥镜图像组成的多中心数据。为 AI-IDPS 开发了 13 个特征提取模型和 1 个特征拟合模型。AI-IDPS 的效率通过 196 张图像和 33 个连续收集的视频进行评估，并与纯深度学习模型和内窥镜医师的表现进行比较。我们进行了交叉研究和问卷调查，以调查该系统对内窥镜医生理解人工智能预测的影响。结果 AI-IDPS 在图像验证中表现出区分 SM2-3 病变的敏感性、特异性和准确性，分别为 85.7%、86.3% 和 86.2%，在连续收集的视频中分别为 87.5%、84% 和 84.9%。纯深度学习模型的敏感性、特异性和准确性均明显较低（分别为 83.7%、52.1% 和 60.0%）。内窥镜医师的准确性（平均从 79.7% 提高到 84.9%，P = 0.03）、敏感性（平均从 37.5% 提高到 55.4%，P = 0.27）和特异性（平均从 93.1% 提高到 94.3%，P = 0.27）显着提高。 = 0.75) AI-IDPS 协助后。讨论 基于领域知识，我们开发了一个可解释的系统来预测 ESCC 侵袭深度。人性疗法展示了在实践中超越深度学习架构的潜力。