Research On Missing Modality Image Classification Based On Machine Learning

Magnetic resonance imaging(MRI)technology can provide high-resolution images of human tissues,occupying an increasingly important position in the medical field.As a type of cancer,brain glioma is the most common malignant tumor in the skull.In the classification of gliomas,because the MRI images of different modalities can reflect different information of the tumor area,doctors often use multiple modalities of brain MRI images as a reference when making judgments to determine the final classification result.Based on the data that has been marked by the doctor,an algorithm for automatically classifying brain tumors can be designed.Unfortunately,in the actual clinical data collection process,many problems will be encountered.One of the most important problems that hinder the use of multi-modal imaging is data loss.For example,the patient has head movement during the filming process,which causes false images.This kind of data will be discarded during screening.In addition,some patients are unwilling to do one of these tests,and some patients' image data will be incomplete.When using multi-modal data to train a model for classification,it is often necessary to discard patient data with missing modalities,causing great waste of data and reducing the number of samples in the analysis.Therefore,how to make better use of observed data to complete the missing data,so as to achieve better classification of brain tumors has become a question of practical significance.At present,for the missing modal problems of medical images,machine learning methods are mainly used to complete the feature data.Based on the machine learning methods,this paper first proposes the incomplete features of multimodal glioma image extraction.A smooth tensor decomposition algorithm based on multi-way delay embedding,and then based on the total variation low-rank completion algorithm,an ensemble learning algorithm based on total variation low-rank completion is proposed,which improves the classification accuracy of brain tumors.The main innovations and contributions of this article are mainly in the following two aspects:(1)A smooth tensor decomposition algorithm based on multi-way delay embedding is proposed.This algorithm improves the decomposition model based on the original multi-way delay-embedded Tucker decomposition algorithm.After the data is fold by the multi-way delay embedding operation,the folded tensor is obtained,and the smoothed tensor CP decomposition is performed on the folded tensor to obtain the completed tensor.Finally,the reverse process of the multi-way delayed embedding is performed.Finally,the reverse process of the multi-directional delayed embedding is performed to obtain complete data.The improved algorithm makes up for the lack of Tucker decomposition of the original algorithm and enhances the denoising ability of the algorithm.(2)An ensemble learning algorithm based on total variation low-rank completion is proposed.By using the total variation low-rank completion algorithm(LRTV),the the multi-way delayed embedding smooth CP decomposition(MDT-SPC),the multi-way delayed embedding tucker decomposition(MDT-Tucker)algorithm completes the feature data to train the base classifiers separately,and uses 5-fold cross-validation to make predictions.Three new features are obtained,and the new feature information is complemented with the data imputed by LRTV algorithm.All the data are combined to achieve better recognition accuracy than traditional methods.