面向自闭症辅助诊断的深度对比模糊神经网络

Deep Contrastive Fuzzy Neural Network for Aided Diagnosis of Autism

静息态功能磁共振成像(rs-fMRI)可有效反映大脑活动状况,然而rs-fMRI数据的高随机性和自闭症谱系障碍(ASD)内在的高异质性给ASD计算机辅助诊断带来了不确定性。提出一种基于对比损失的Takagi-Sugeno-Kang(TSK)深度模糊神经网络CL-DeepTSK,结合多输出TSK(MO-TSK)模糊系统与多层感知机(MLP)有效缓解数据不确定性对模型的影响,提升TSK模糊系统的表达能力,并使模型更具可解释性。使用对比损失目标学习准则对MO-TSK与MLP进行联合优化,提高训练样本缺乏时的模型泛化性能。在ABIDE数据集上的实验结果表明,CLDeepTSK的平均正确率和AUC指标分别达到70.0%和0.773,同时获得了30个最具鉴别性的功能连接。上述实验结果证明了CL-DeepTSK能够有效地进行自闭症辅助诊断,并且具有较高的可解释性。

Resting-state functional Magnetic Resonance Imaging(rs-fMRI)can effectively reflect brain activity.However,the high randomness in rs-fMRI data and high heterogeneity in autism cases cause high uncertainty in the diagnosis of Autism Spectrum Disorder(ASD).Hence,this study integrates a fuzzy system with a deep neural network and proposes a Takagi-Sugeno-Kang(TSK)deep fuzzy neural network based on Comparative Loss(CL),which is abbreviated as CL-DeepTSK.CL-DeepTSK combines the Multi-Output TSK(MO-TSK)fuzzy system with a Multilayer Perceptron(MLP),which effectively reduces the effect of data uncertainty on the model,improves the expression ability of the TSK fuzzy system,and renders the model interpretable.Additionally,MO-TSK and MLP are jointly optimized using a novel CL objective-learning criterion,which improves the generalization performance of the model when the training samples are insufficient.For the Autism Brain Imaging Data Exchange(ABIDE)dataset,the accuracy and Area Under Curve(AUC)of the CL-DeepTSK are 70.0%and 0.77,respectively,and the 30 most discriminative functional connections are obtained.Experimental results show that the proposed CL-DeepTSK model can be effective and interpretable for the auxiliary diagnosis of ASD.