面向肺结节多语义特征分类的不确定性多任务损失方法

肺结节的早期诊断对后续的治疗非常重要。 尽管深度学习方法在肺结节良恶性分类等任务中取得了良好的结果,但是这些方法没有提供有意义的诊断功能,导致获得的结果缺乏客观性。越来越多的研究者引入了肺结节的其他语义特征来解决这个问题,但是多个语义特征的引入会造成模型的负迁移。为了解决肺结节多个语义特征之间同步共享知识的程度不 同造成的负迁移问题。本文提出一种基于不确定性多任务损失的深度学习模型,对肺结节的9个语义特征(精细度,内部结构,钙化,球形度,边缘,分叶征,毛刺征,纹理,恶性程度)进行分类,通过每个任务的同质不确定性来权衡多个损失函数的权重。我们在基准数据集LIDC-IDRI上验证了该方法,本文提出的模型在恶性程度上的分类准确率为93.6%,ROC曲线下面积为95.5%,查全率为84.6%,特异性为94.5%。我们的模型通过肺结节多个语义特征共享知识的程度不同进而改变多个语义特征相对权重提高了恶性程度的分类性能。

Uncertain multi-task loss method for multi-semantic feature classification of lung nodules

It is important to diagnose lung nodules as early as possible for sub sequent treatment.Although deep learning methods have achieved good results in tasks such as classification of malignant tumors,these methods do not provide m eaningful diagnostic functions resulting in the results obtained lack objectiven ess.More and more researchers have introduced other semantic features of lung n odules to solve this problem,but the introduction of multiple semantic features will cause negative migration of the model.In order to solve the problem of ne gative transfer caused by the different degrees of synchronized sharing of knowl edge between multiple semantic features of lung nodules.This paper proposes a d eep learning model based on uncertain multitask loss,classifies nine semantic f eatures of lung nodules (subtlety,internal structure,calcification,sphericity ,margin,lobulation,spiculation,texture and malignancy),and weighs multiple loss functions through the homogenous uncertainty of each task.We verified the method on the benchmark LIDC-IDRI dataset.The model proposed in this paper ha s a classification accuracy of 93.6% in malignancy,an area under the ROC curve o f 95.5%,a sensitivity of 84.6%,and a specificity of 94.5%.Our model improves the classification performance of malignancy by changing the relative weight of multiple semantic features through different semantic features of lung nodules.