本研究旨在开发一个能够自动定位息肉并按NICE分型的深度学习目标检测模型，以促进更有效的诊断和治疗规划。

收集2018年1月至2023年6月来自苏州大学附属常熟医院、常熟市中医院和常熟市辛庄人民医院的4个结肠息肉数据集，包括静态图像和视频。所有样本经病理学证实并按NICE分型分类。图像使用LabelMe工具进行标注，随后转换成MSCOCO格式以适配深度学习模型训练。采用预训练的Faster R-CNN模型，结合实时数据增强和多种图像处理技术，通过迁移学习策略进行模型训练。模型的性能评估遵循COCO标准，关注交并比（IoU）、平均精度（AP）和召回率。此外，对模型在NICE分型中的识别能力进行了详细评估。

分析1 835例患者的2 248个结直肠息肉，所有息肉经病理学证实，并按NICE分型分类，具体为NICE 1型575例，NICE 2型1 143例，和NICE 3型530例。通过迁移学习技术，开发了基于ResNet-50和ResNet-101骨干网络的Faster R-CNN模型，其中以ResNet-101为基础的版本被命名为Faster R-CNN-NICE。性能分析显示，虽然Faster R-CNN-NICE模型在处理速度上略有下降（减少1.72帧/秒），但在边界框平均精度（bbox_mAP达0.542）和息肉分类综合平均精度（mAP为0.830）上显著提升。与内镜医生相比，此模型在大部分情况下展示了更高的置信度和准确性，特别是在NICE 2型息肉的预测中，其性能与医生的预测结果的差异存在统计学意义（χ²=4.30，P<0.05）。模型转换为ONNX格式后，在不同硬件环境下显示了优化的执行效率，并在视频实时检测中有效地实现了息肉的快速定位和精确分类。

本研究根据NICE分型系统构建了一个结肠息肉图像数据集，并开发出Faster R-CNN-NICE深度学习模型，有效地实现了结肠息肉的即时定位与精确鉴别。这项技术预期将为内镜医生的临床工作提供支持，显著提升诊断的效率和准确度。

This study aims to develop a deep learning object detection model capable of autonomously locating polyps and classifying them according to the NICE classification, thereby facilitating more effective diagnostic and treatment planning.

This study utilized four colorectal polyp datasets collected from Changshu Hospital Affiliated to Soochow University, Changshu Traditional Chinese Medicine Hospital and Changshu Xinzhuang People's Hospital from January 2018 to June 2023, comprising both static images and videos. All samples were pathologically confirmed and classified according to the NICE typology. The images were annotated using the LabelMe tool and subsequently converted to MSCOCO format to suit deep learning model training. We employed a pre-trained Faster R-CNN model, combined with real-time data augmentation and various image processing techniques, and trained the model using a transfer learning strategy. Model performance was evaluated following the COCO standards, focusing on Intersection over Union (IoU), Average Precision (AP), and recall rate. Additionally, the model's capability in NICE classification recognition was thoroughly assessed.

We analyzed 2 248 colorectal polyps from 1 835 patients, all pathologically confirmed and classified as 575 NICE type 1, 1 143 NICE type 2, and 530 NICE type 3. Using transfer learning techniques, Faster R-CNN models based on ResNet-50 and ResNet-101 backbone networks were developed, with the ResNet-101-based version named Faster R-CNN-NICE. Performance analysis showed that although the Faster R-CNN-NICE model experienced a slight decrease in processing speed (reduction by 1.72 frames per second), it significantly improved in bounding box average precision (bbox_mAP reaching 0.542) and comprehensive average precision for polyp classification (mAP at 0.830). Compared to endoscopists, this model demonstrated higher confidence and accuracy in most cases, particularly in predicting NICE type 2 polyps, where its performance significantly differed from that of the doctors (χ²=4.30, P<0.05). After conversion to the ONNX format, the model exhibited optimized execution efficiency in various hardware environments and effectively achieved rapid localization and accurate classification of polyps in real-time video detection.

This research constructed a colon polyp image dataset based on the NICE classification system and developed a Faster R-CNN-NICE deep learning model, which effectively achieves immediate localization and precise identification of colon polyps. This technology is expected to support the clinical work of endoscopists, significantly enhancing the efficiency and accuracy of diagnoses.