探讨基于深度学习的结直肠息肉诊断自动分类模型的构建。

收集2018年1月至2023年1月在苏州市3个内镜中心的不同图像增强内镜(IEE)技术下的结肠镜图像957张(常熟市第一人民医院537张，常熟市中医院359张，苏州大学附属第一医院61张)，依据病理结果分为正常组、增生性息肉组和腺瘤性息肉组。利用DenseNet-121、EfficientNet、resnet101和resnet50 4种卷积神经网络(CNN)框架，构建深度学习模型，并评估其与经验不同的内镜医师的准确率、召回率、精确度、F1值和读片时间。

EfficientNet在4个模型中最为优越，准确率0.961，召回率0.968，精确度0.959，F1值0.962，在读图用时方面，所有模型完成图像自动诊断任务的平均时间为(4.08±0.63)s，远快于内镜医师所需的平均时间[(291.10±17.68)s]，差异有统计学意义(t＝－36.22，P<0.01)。将EfficientNet预训练模型经迁移学习后的模型命名为"EffiPolyNet"，其在腺瘤性息肉上有少量误分类，但准确率达0.90，AUC为0.98。t-分布随机邻域嵌入(t-SNE)可视化揭示了腺瘤性和增生性息肉间部分语义特征重叠，解释了模型的误分类。利用梯度加权分类激活映射(Grad-CAM)和沙普利可加性解释(SHAP)，揭示了模型决策中的关键图像区域和特征的相对重要性。

EffiPolypNet模型在多种IEE技术的结直肠息肉性质分类中表现出色，为结肠镜光学诊断提供了高效且可靠的支持。

To explore the construction of automatic classification model for the diagnosis of colorectal polyp based on deep learning.

From January 2018 to January 2023, 957 colonoscopy images were collected at 3 endoscopy centers in Suzhou (537 at Changshu NO.1 People′s Hospital, 359 at Changshu Hospital of Traditional Chinese Medicine, and 61 at the First Affiliated Hospital of Soochow University), by using various image enhanced endoscopy (IEE) techniques. Based on pathological features, these images were classified into normal group, hyperplastic polyps group, and adenomatous polyps group. By using the DenseNet-121, EfficientNet, ResNet101, and ResNet50 convolutional neural network (CNN) frameworks, deep learning models were constructed and tested against the performance of endoscopists with varied experience levels in terms of accuracy, recall rate, precision, F1 score, and time taken to read images.

EfficientNet outperformed the other models, with a 0.961 accuracy, 0.968 recall rate, 0.959 precision, and 0.962 F1 score. In terms of image reading time, all models significantly outperformed endoscopists, completing automatic diagnostic tasks in an average time of (4.08±0.63) seconds compared to the average time of (291.10±17.68) seconds required by endoscopists, showing a statistical difference (t=-36.22, P<0.01). The EfficientNet pretrained model, after transfer learning, was named " EffiPolypNet". It misclassified a few adenomatous polyps but achieved an accuracy of 0.90 and an AUC of 0.98. Visualization using t-distributed stochastic neighbor embedding (t-SNE) revealed semantic feature overlaps between adenomatous and hyperplastic polyps, which could account for the misclassifications. Gradient-weighted class activation mapping (Grad-CAM) and Shapley additive explanations (SHAP) elucidated the key image regions and relative importance of features in the model′s decision-making process.

The EffiPolypNet model outperformed other IEE techniques in categorizing the nature of colorectal polyps, offering efficient and dependable support for optical diagnosis in colonoscopy.