基于元学习和结肠镜多模态图像的NBI结肠镜辅助诊断模型研究

基于深度学习的NBI结肠镜辅助诊断模型对结肠病变的分类识别目前达到了很高的准确度，但由于对结肠病变类型样本分布覆盖不全面，并且多种病变类型的样本数量有限，使得已有研究中的图像训练集在细粒度的众多结肠病变类型完整分布上表现为一个少样本子集。对这种少样本子集数据的过拟合，导致已有辅助诊断模型大都集中在对几类常见结肠病变的简单粗粒度分类，临床通用性不强。本项目拟改进深度网络模型的训练方式，采用元学习方法，利用结肠镜多模态图像组建大量少样本分类任务迭代训练网络模型，让模型学会在少量样本上分类，从而降低对训练数据量的依赖，减少训练集样本分布的影响，进而解决NBI结肠镜图像的少样本分类问题，建立对真实分布的各种常见和罕见及未来出现的结肠病变分类识别的NBI结肠镜辅助诊断模型。采用元学习解决结肠镜影像分析领域的少样本学习问题之前未见报道，项目的研究将为推进结肠镜计算机辅助诊断技术的临床应用提供支持证据。

The existing NBI colonoscopy computer-aided diagnosis models based on deep learning are now achieved a high degree of accuracy in classification of colonic lesions. However, due to the incomplete coverage of the sample distribution of colon lesions and the limited number of samples of various lesions, NBI colonoscopy image training set in previous researches appears as a small sample subset over the complete distribution of a wide variety of fine-grained colonic lesion types. Deep learning tends to overfit on this kind of few samples, which makes most studies in the current literature have applied simple coarse-grained categories in differentiating the type of colon polyp and concentrated on common colon lesions in several categories. Thus, the existing auxiliary diagnosis system has low adaptability and versatility in clinical application. This project intends to improve the training method of the model. Meta-learning has been proposed as a framework to address the challenging few-shot learning setting. Leverage a large number of few-shot classification tasks built on multi-modality colonoscopy images to train deep neural networks episodically in order to let the model learn to classify on a few samples, thereby overcoming the necessity of large amount of data and reducing the impact of the distribution of training samples on the model performance. Finally, solve the small sample size problem on NBI colonoscopy image classification and build an NBI colonoscopy computer-aided diagnosis model to achieve classification of various common and rare types of colonic lesions with real distribution and future colonic lesions. Using meta-learning to solve few-shot learning in the field of colonoscopy image analysis has not been reported previously. Our research will provide supporting evidence for promoting the implementation of computer-aided diagnosis technology into routine colonoscopy.