本文選題：功能性磁共振成像(fMRI)　切入點(diǎn)：過擬合　出處：《模式識別與人工智能》2016年07期

【摘要】：針對功能性磁共振成像(f MRI)數(shù)據(jù)高維小樣本特性給分類模型帶來的過擬合問題,文中基于Softmax回歸提出結(jié)合L2正則與L1正則的全腦f MRI數(shù)據(jù)特征選擇框架.首先,基于大腦認(rèn)知的特點(diǎn),將全腦分成感興趣區(qū)域和非感興趣區(qū)域.然后,使用可以縮小權(quán)值系數(shù)的L2正則對感興趣區(qū)域建模以選出感興趣區(qū)域的全部體素,使用具有稀疏作用的L1正則對非感興趣區(qū)域建模以選出非感興趣區(qū)域中的激活體素.最后,結(jié)合感興趣區(qū)域和非感興趣區(qū)域的體素構(gòu)成全腦f MRI數(shù)據(jù)的正則化Softmax回歸模型.在Haxby數(shù)據(jù)集上的實驗表明,L2與L1的正則化策略可有效提升全腦分類的準(zhǔn)確率.
[Abstract]:In order to solve the problem of over-fitting of the classification model caused by the high dimensional and small sample characteristics of functional magnetic resonance imaging (fMRI) data, a framework for feature selection of global brain f MRI data combining L2 canonical and L1 regularization is proposed based on Softmax regression.First, the whole brain is divided into regions of interest and non-regions of interest based on the cognitive characteristics of the brain.Then, the region of interest is modeled by L _ 2 canonical which can reduce the weight coefficient to select all voxels of the region of interest, and the active voxels in the region of interest are obtained by using L _ 1 canonical model with sparse function.Finally, the regularized Softmax regression model of the global brain f MRI data is constructed by combining the voxels of the region of interest and the region of interest.Experiments on Haxby datasets show that the regularization strategies of L2 and L1 can effectively improve the accuracy of global classification.
【作者單位】： 北京工業(yè)大學(xué)計算機(jī)學(xué)院多媒體與智能軟件技術(shù)北京市重點(diǎn)實驗室;首都醫(yī)科大學(xué)宣武醫(yī)院;
【基金】：國家重點(diǎn)基礎(chǔ)研究發(fā)展計劃(973計劃)項目(No.2014CB744601) 國家自然科學(xué)基金項目(No.61375059;61332016) 北大方正集團(tuán)有限公司數(shù)字出版技術(shù)國家重點(diǎn)實驗室開放課題資助~~
【分類號】：R445.2;O482.532
，

本文編號：1721693