Research on automatic classification of breast MRI images based on deep learning
Objective:To train and verify an automatic image classification model based on deep learning and program.Materials and Methods:A total of 862 breast MRI images were collected from the picture archiving and communication system(PACS)system from January 2010 to November 2020(dataset Ⅰ).The images were divided into three categories:T2WI,fat-suppressed(FS)T2WI,apparent diffusion coefficient(ADC).A deep learning model of sequence differentiation was trained with the dataset Ⅰ.Another group of 377 breast MRI images from February 2013 to April 2020(data set Ⅱ)were collected and divided into three categories:no-contrast(NoC),contrast enhanced early(CEearly),contrast enhanced(CE)according to the phase characteristics of dynamic contrast-enhanced(DCE).A deep learning model of phase differentiation of DCE was trained with the dataset Ⅱ.A third group of 95 breast MRI images(data set Ⅲ)were collected from October 2021 to December 2021 for independent validation of the classification models(different sequences and different phases of DCE).Then the diffusion weighted imaging(DWI)parameters were classified by the program in the data set Ⅲ(DWI-high and DWI-low).Using the classification results of radiologists on images as the gold standard,according to image sequence,enhancement characteristics and parameters on as the gold standard,the confusion matrix was used to evaluate the classification performance of the model.Results:In the sequence classification model,the overall prediction accuracy was 92.0％,and the prediction accuracy for each sequence of ADC,T2WI and FS T2WI were 100.0％,84.9％,and 100.0％,respectively.In the DCE classification model,the overall prediction accuracy was 90.4％,and the prediction accuracy for each sequence of NoC,CEearly,and CE were 89.7％,39.2％,and 95.7％,respectively.The program's classification of DWI-high and DWI-low was exactly the same as that of the radiologist.Conclusions:Using deep learning model and program technology to classify the image sequence,phase and parameters of multi-parameter breast MRI,the output results are highly consistent with the classification results of physicians,which basically meet the clinical needs.
breast tumorautomated classification of imagesdeep learningartificial intelligencemagnetic resonance imaging
万方数据
维普
