Application Of Artificial Intelligence In MRI-assisted Diagnosis Of Tuberous Sclerosis Complex In Children

Tuberous sclerosis complex（TSC）is an autosomal dominant genetic disorder caused by mutations in the TSC1 or TSC2 gene,and epilepsy caused by it is one of the symptoms that most affects the quality of life.The main treatment for epilepsy is antiepileptic drugs,but many patients are drug-resistant and can only be found after taking the drugs for a period of time.In addition,some lesions,cortical tubers in patients with TSC are epileptogenic foci,and their location and morphology are important for the diagnosis and treatment of epilepsy.Therefore,it is very important to predict the patient’s outcome as early as possible according to the patient’s symptoms,and to detection and localization of cortical tuber lesions of TSC patients on magnetic resonance image（MRI）as soon as possible.At present,machine learning and deep learning methods have achieved good results in data analysis and outcome prediction of various diseases including epilepsy.The great achievements of deep learning in the field of computer vision have also been well applied to the classification,localization,and segmentation of medical images.However,both machine learning and deep learning require a large amount of data.For TSC,a rare disease,related research is difficult to obtain convincing results due to the small amount of data and the unclear clinical mechanism.Therefore,through the cooperation with Shenzhen Children’s Hospital,a database of children with TSC was constructed.The clinical data of 103 children with TSC were collected,and the multicontrast magnetic resonance（MR）image data of 95 children with TSC were collected,and the TSC lesions were annotated by radiologist at the pixel level.Based on this,this paper proposes a machine learning based TSC epilepsy outcome prediction method and a deep learning based automatic segmentation method of TSC lesions.First,this paper introduces the current common machine learning algorithm models and the application of deep learning models in the image field to provide background knowledge for TSC epilepsy outcome prediction and automatic segmentation of TSC lesions.Second,this paper proposed a set of methods for predicting outcomes of TSC patients based on machine learning models.First,by comparing different combinations of 3 common feature selection methods（analysis of variance F-test,chi-square coefficient,mutual information）and 6 machine learning models（decision tree,random forest,support vector machine,naive Bayes,logistic regression,multi-layer perceptron）,the combination with the highest classification performance was found.Then its parameters were adjusted to be optimal for the epilepsy outcome prediction task.Finally,using permutation importance and a novel method based on it to evaluate the importance of each and every part of the feature,and using another novel method,by gradually removing information and comparing the reduction of the performance,to explore which aspects of information of the MRI features is more important,the more important features and aspects of information for the prediction model were found.Then,this paper proposed a novel multi-modal fusion network and a novel multiview fusion network for TSC lesion segmentation based on U-Net,which can improve the segmentation performance with multi-modal MR image data and with multi-view MR image data,respectively.The multi-modal fusion network fused multiple modalities in the U-Net encoder and added a spatial attention mechanism,while the multi-view fusion network used two U-Net encoders to encode images of two views respectively,and then they were fused in the designed fusion module,and finally the decoder output the segmentation result.The results showed that the Dice coefficient of the multi-modal fusion network with multi-modal data was improved by 11% compared to the single-modal U-Net,and the multi-view fusion network with multi-view data improved the Dice coefficient by 4% compared to the single-view U-Net.In summary,this paper is dedicated to applying machine learning and deep learning methods to TSC outcome prediction and automatic segmentation of TSC lesions,and strives to improve the performance of the model and achieved good results.