基于特征融合的卷积神经网络乳腺癌图像分类

传统乳腺癌图像分类方法需要从医学图像中人工提取特征，不仅需要具备专业医学知识，而且存在耗 时费力、提取高质量特征困难等问题.因此，提出了一种基于特征融合的卷积神经网络乳腺癌图像分类方法. 首先预训练了两个不同结构的卷积神经网络，然后利用卷积神经网络自动提取特征的特性，将两个结构提取 到的特征进行融合，最后利用分类器对融合的特征进行分类；同时，为避免卷积神经网络模型受小样本量限 制出现过拟合现象，通过乳腺病变区域提取、区域细化和数据增强等方法对图像进行适当预处理，并通过过 采样方法解决了正负样本不平衡的问题.实验结果显示，该方法在乳腺癌图像数据集BCDR-F03上分类AUC达 到89%，对乳腺癌图像的分类精度较传统方法有明显提高.
Traditionalbreastcancerimageclassificationmethodsneedtoextractfeaturesfrommedicalimages,which notonlyrequiresprofessionalmedicalknowledge,butalsoistime-consuming,laboriousanddifficulttoextracthighqualityfeatures.Therefore,aconvolutionneuralnetworkbasedonfeaturefusionisproposedforbreastcancerimageclas? sification.Thefirstpre-trainingoftwodifferentstructureoftheconvolutionalneuralnetwork,andthenusetheconvolu? tionneuralnetworkcharacteristicsofautomaticfeatureextraction,featureextractiontothetwostructuralintegration,fi? nallyusingclassifiertoclassifythefeatureoffusion;atthesametime,inordertoavoidtheconvolutionalneuralnetwork modelbythesmallsamplesizelimitoverfittingthisphenomenon,throughregionalextraction,refinementandenhance? mentofbreastlesionsofregionaldataandothermethodsofproperimagepretreatment,andthesamplingmethodtosolve theproblemoftheimbalanceofpositiveandnegativesamples.Theexperimentalresultsshowthatthismethodcanclassi? fyAUCinto89%onthedatasetBCDR-F03,andtheclassificationaccuracyofbreastcancerimagesissignificantlyim? provedcomparedwiththetraditionalmethods.