基于多时期蒸馏网络的随访数据知识提取方法

Knowledge extraction method for follow-up data based on multi-term distillation network

随着医学上对随访工作的不断重视,通过医学图像分析的方法获取随访指导的相关信息变得越来越重要;然而,在深度学习领域,大多数方法不适用于处理此类任务。为了解决这个问题,提出了一种多时期知识蒸馏(MKD)模型。首先,借助知识蒸馏在模型迁移方向上的优势,将带有长时期随访信息的分类任务转换为基于领域知识的模型迁移任务;然后,充分利用长时期医学图像中所包含的随访知识,来完成长时期的肺结节分类。同时,针对随访过程收集到的数据每一年相对不平衡的问题,提出了一种基于元学习思想的正则化方法,且该方法能够有效地在半监督模式下提高模型的训练精度。在NLST数据集上的实验结果表明,所提出的MKD模型在长时期肺结节分类任务下较GoogleNet等深度学习分类模型的分类精度更优越。在不平衡数据量达到800例时,利用元学习方法改进后的MKD模型相较于现有先进模型最高可以提升7个百分点的精度。

As medical follow-up work is more and more valued,the task of obtaining information related to the follow-up guidance through medical image analysis has become increasingly important.However,most deep learning-based methods are not suitable for dealing with such task.In order to solve the problem,a Multi-term Knowledge Distillation(MKD)model was proposed.Firstly,with the advantage of knowledge distillation in model transfer,the classification task with long-term follow-up information was converted into a model transfer task based on domain knowledge.Then,the follow-up knowledge contained in the long-term medical images was fully utilized to realize the long-term classification of lung nodules.At the same time,facing the problem that the data collected during the follow-up process were relatively unbalanced every year,a meta-learning method based normalization method was proposed,and therefore improving the training accuracy of the model in the semi-supervised mode effectively.Experimental results on NLST dataset show that,the proposed MKD model has better classification accuracy in the task of long-term lung nodule classification than the deep learning classification models such as GoogleNet.When the amount of unbalanced long-term data reaches 800 cases,the MKD enhanced by meta-learning method can improve the accuracy by up to 7 percentage points compared with the existing state-of-the-art models.