Prediction Of Pathological Grade In NF-pNETS Based On Multimodal Imaging

Pancreatic neuroendocrine tumors are highly heterogeneous tumors that seriously threaten human health,of which about 70% are non-functional pancreatic neuroendocrine tumors(NFp NETS).For NF-p NETS patients with higher grades of malignancy,surgical resection is the most effective treatment in clinical practice;for patients with lower grades of malignancy,drug therapy is more appropriate.Therefore,accurate grading is important to precise clinical treatment of NF-p NETS.At present,the clinical determination of NF-p NETS grade is mainly based on the visual assessment of unimodal imaging data,and the grade of the tumor is given by the clinician's personal experience and subjective observation and analysis.However,the visual assessment method is not only time-consuming and labor-intensive,but also has strong personal subjectivity,and the accuracy of predicting the pathological grade of NF-p NETS using unimodal imaging data is poor.The purpose of this paper is to analyze the CT and MRI modal images of patients and build a grading model to achieve accurate prediction of the pathological grade of NF-p NETS and provide objective support for clinicians' diagnosis.CT modal images include two sequences of arterial phase and delayed phase images;MRI modal images include two sequences of T1-enhanced and T2-enhanced images.In this paper,a radiomics model for pathological grading of NF-p NETS was first constructed.In order to increase the amount of data available for analysis in the experiment,Elastic software was first used to achieve registration between different sequences of images for each modality.Then,the radiomic features of the tumor region were extracted from the CT and MRI modal images respectively,and the multi-modal features were fused using the fusion methods of feature concatenation and CCA.Finally,the optimal feature subset is selected using rank sum test,m RMR and LASSO algorithm.In terms of experimental design,this paper uses SVM and LDA classifiers to construct different NF-p NETS grade prediction models according to different optimal feature subsets.The experimental results show that the prediction model based on CCA fusion of CT and MRI modal features(CCA-model)built using the LDA classifier achieves the best prediction results on the training set and test set,with AUC values of 0.983 and 0.853,respectively.In order to solve the problem that the radiomics model needs to be based on artificially designed image features,this paper further proposes a multi-branch fusion convolutional neural network for NF-p NETS pathological grading.The network uses a different sequence of images for each modality to achieve the rank prediction of NF-p NETS,and consists of two branches,each responsible for processing a sequence of images.The network structure uses a parameterfree and adaptive feature layer replacement fusion method,which exchanges feature layers between different network branches.That is,the process of feature layer replacement is guided by the importance of each feature layer itself.Determined by the size of the scale factor of the BN layer during network training.At the same time,all parameters except the BN layer are shared among different network branches,which makes the parameter amount and structure of the network as compact as a single branch network.Experiments show that the multi-branch fusion network structure proposed in this work achieves better NF-p NETS classification accuracy than the single-modal model.The F1 scores of the model trained with the two CT sequences on the training set and the test set reached 0.828 and 0.713,respectively;the F1 scores of the model trained with the two MRI sequences on the training set and the test set reached 0.824 and 0.746,respectively.