| Contrast-enhanced imaging,which can produce images with higher contrast and sharpness.It’s widely used in clinical diagnosis of various diseases.Multiple sets of images taken within a certain time interval after the contrast agent’s injection are called multi-phase images.Multi-phase imaging improves the specificity and sensitivity of diagnosis,but the lack of experience and ability often makes it difficult for junior doctors to use multi-phase images in diagnosing.At present,artificial intelligence methods have been used to assist multi-phase imaging diagnostic research due to their ease of use and high compatibility with medical images,but most of the current research can only directly give diagnosis results without helping training junior doctors.In order to assist the training of junior doctors in multi-phase diagnosis and improve the interpretability of artificial intelligence methods in computer-aided diagnosis(CAD),a machine learning multiphase image-assisted diagnosis model construction experiment using traditional radiomics methods was carried out at first.In this experiment,dynamic contrast-enhanced MRI(DCE-MRI)data was used to predictβ-catenin mutations in patients with Hepatocellular carcinoma(HCC).Afterwards,this research proposed an interpretable neural network with the radiomics features of multiphase images as input based on the multi-tower model structure,which was improved by the neural network two-tower model.The four-period Dynamic Contrast-Enhanced CT(DCE-CT)was used as the material to carry out HCC differentiation prediction experiment.Multi-tower model can independently process radiomics features from different phases in the input branch,and fuse deep features from independent input branches at the top of the model to finally predict the target.After training,the weights are extracted from the nodes in the interoperability stage.The importance of images is calculated as a prompt for diagnosis and fed back to the radiologist.Finally,a modelassisted reading experiment was carried out to verify whether it had a practical diagnosing aiding ability.In the radiomics model application experiment,the combination of radiomics features from the four phases DCE-MRI was applied in different models,and the Hepato-biliary Phase(HBP)single-phase model had the best test set results,and the area under receiver operating characteristic curve(ROC-AUC)reached 0.82(95%CI:0.63-0.93);Most of the prediction performance indicators obtained by the multitower model in the model experiment were higher than those of the traditional radiomics machine learning model and the fusion machine learning model.The ROCAUC(hereinafter referred to as AUC)in the test set was 0.85(95%CI:0.77-0.92),the accuracy(ACC)was 0.63,the sensitivity(SEN)was 0.76,specificity(SPE)was 0.84,and the results were significantly different from the next highest performance model by the Delong test.With the help of image importance ranking,the average SEN,SPE,ACC and AUC of the four doctors participating in the reading experiment increased by 0.11,0.08,0.10 and 0.09,respectively,which indicated that the image importance ranking composed of model node weights could effectively improve the doctors’diagnosing ability.In this study,a CAD application experiment with radiomics was performed,and a multi-phase imaging diagnostic support model based on radiomics and neural network was proposed.The experimental results showed that the predictive performance of the model was higher than that of the conventional radiomics model.The image importance ranking from model could effectively improve the ability of radiologists to use multiphase phase in diagnosis. |
