基于深度学习的人工智能模型在窄带光及近焦窄带光内镜图像中精确识别早期胃癌边界的临床研究

A clinical study of deep learning‑based artificial intelligence model for precise identification of early gastric cancer boundaries in narrow‑band and near focus narrow‑band endoscopic images

摘要:

目的 开发基于深度学习的人工智能模型用于窄带光成像（narrow‑band imaging，NBI）及近焦窄带光成像（near focus narrow‑band imaging，NF‑NBI）图像中识别早期胃癌的边界，并评估模型识别边界的能力。方法 回顾性收集2016年2月至2024年6月于苏州大学附属第一医院消化内科行内镜黏膜下剥离术（endoscopic submucosal dissection，ESD），术后病理确诊为早期胃癌的282例患者的内镜图像。将图像按近似8∶2的比例随机划分为训练集及验证集。在NBI模态中，171例患者的980张图像用于训练，61例患者的235张图片用于验证；在NF‑NBI模态中，128例患者的1 273张图像用于训练，35例患者的373张图像用于验证。共训练6个卷积神经网络（convolutional neural network，CNN）模型，包括2个独立的CNN1模型、2个独立的CNN2模型和2个融合的CNN3模型。以内镜专家参考ESD术后病理复原图描绘的早期胃癌边界为金标准，在验证数据集中比较CNN3模型与低年资内镜医师（5年以下操作经验）、中年资内镜医师（5~10年操作经验）、高年资内镜医师（10年以上操作经验）识别早期胃癌边界的交并比等。结果 在NBI验证集图像上CNN3模型的交并比值显著高于低年资内镜医师（0.732 比 0.489，Z=11.528，P<0.001）以及中年资内镜医师（0.732 比 0.521，Z=11.184，P<0.001），但CNN3模型与高年资内镜医师之间差异无统计学意义（0.732 比 0.739，Z=0.593，P=0.554）。在NF‑NBI验证集图像上CNN3模型的交并比值显著高于低年资内镜医师（0.757 比 0.537，Z=15.944，P<0.001）以及中年资内镜医师（0.757 比 0.597，Z=9.722，P<0.001），但CNN3模型与高年资内镜医师之间差异无统计学意义（0.757 比 0.769，Z=0.854，P=0.394）。结论 融合模型CNN3在NBI图像和NF‑NBI图像上均体现出接近于高年资内镜医师对早期胃癌边界识别的能力，该模型可以有效协助中、低年资内镜医师精确识别早期胃癌边界。

Abstract:

Objective To develop and validate artificial intelligence (AI) models based on deep learning for precise boundary identification of early gastric cancer (EGC) in narrow‑band imaging (NBI) and near focus narrow‑band imaging (NF-NBI) endoscopic images. Methods Endoscopic submucosal dissection (ESD) images from 282 patients diagnosed as having EGC by postoperative pathology at the Department of Gastroenterology, the First Affiliated Hospital of Soochow University were retrospectively collected from February 2016 to June 2024. The images were randomly divided into the training set and the validation set at an approximate 8∶2 ratio. In the NBI modality, 980 images from 171 patients were used for training, 235 images from 61 patients were used for validation. In the NF-NBI modality, 1 273 images from 128 patients were used for training, and 373 images from 35 patients were used for validation. This study trained a total of six convolutional neural network (CNN) models: two independent CNN1 models, two independent CNN2 models, and two fused CNN3 models. Using expert-delineated EGC boundaries based on post-ESD pathological findings as the gold standard, the intersection over union (IOU) value of the CNN3 models was compared against junior (<5 years experience), mid-level (5-10 years), and senior (>10 years) endoscopists. Results In NBI validation set, the IOU value of CNN3 model was significantly higher than that of junior (0.732 VS 0.489, Z=11.528, P<0.001) and mid-level endoscopists (0.732 VS 0.521, Z=11.184, P<0.001). However, no significant difference was observed between CNN3 model and senior endoscopists (0.732 VS 0.739, Z=0.593, P=0.554). Similarly, in NF-NBI validation set, CNN3 model outperformed junior (0.757 VS 0.537, Z=15.944, P<0.001) and mid-level endoscopists (0.757 VS 0.597, Z=9.722, P<0.001), while matching senior endoscopists (0.757 VS 0.769, Z=0.854, P=0.394). Conclusion The fused CNN3 model achieves senior expert-level accuracy in delineating EGC boundaries in both NBI and NF-NBI images, demonstrating potential to assist less-experienced endoscopists in precise identification of EGC boundaries.