Study On The Correlation Between Brain Plasticity And Prognosis In Hypoxic-ischemic Brain Damage Assessed By Radiomics

Objective: This study systematically studied the scientific diagnosis of hypoxic-ischemic encephalopathy(HIE)by using deep learning and radiomics.By collecting cases of children with HIE,standardizing information of HIE children,and classifying different sequence image lesions of MRI resonance,the MR image lesion database was established for HIE children.Based on u-NET and other deep neural network models,automatic segmentation methods of different types of lesions in HIE brain parenchyma were studied.The features of MRI lesions with different sequences were extracted and a database model was established.The U-NET model combined with residual neural network Res Net18 was used to train the artificial neural network with the morphological data of brain MRI lesions manually marked by radiologists,and an automatic segmentation model was established to accurately segment the lesions in the brain region of HIE children.At the same time for brain development in children with HIE situation to carry on the quantitative evaluation,combined with a series of images of omics characteristics with the traditional NBNA score index,establish neural plasticity in patients with HIE statistical models to forecast,put forward a combination of deep learning,image omics and clinical comprehensive evaluation method of HIE prognosis evaluation system.It provides theoretical basis and research support for scientific diagnosis,treatment and prediction of brain plasticity of HIE children.Methods: 1.A total of 309 neonates diagnosed with HIE in several hospitals from April2019 to August 2021 were selected,the entry criteria were based for the diagnosis of HIE in 2005,and the database of MRI information learned by deep learning model and omics characteristics of children with HIE.Using the lesion segmentation module of Deep Wise Medical scientific research platform,A Res Net + U-net deep learning model was trained to automatically classify lesion in the incoming MRI images.The model was trained with color enhancement modules(Mirroring,Elastic Deformations,Rotation,Scaling,Resampling),noise enhancement module(Gaussian Noise,Rician Noise),tailoring module(Random Crop,and Center Crop).Finally,Deep Wise Medical scientific research platform was used for statistical analysis.2.The children were scored by NBNA,and magnetic resonance examination data from 3to 10 days after birth(after correcting gestational age of premature infants at full term)were taken.The sample size was based on the previous study,and 309 HIE newborns were included into the database.The histomorphological analysis was conducted in terms of the number of lesions in T1 WI sequence,T2 WI sequence,DWI sequence and SWI sequence.For each child,two categories(0,1)tagged,0 represents NBNA score 35 and 1represents NBNA score <35.Magnetic resonance images of all 309 children were uploaded to the Deep Wise online platform.ROI regions of interest were labeled on the platform.To evaluate the prediction efficiency of each model combined with NBNA score: AUC > 0.9 has high accuracy;0.7 < AUC < 0.9 is moderate accuracy;0.5 < AUC< 0.7 is low accuracy;AUC < 0.5 has no diagnostic value.P < 0.05 was considered as a statistically significant difference.A De Long test was done for the model and ROC curves was performed to compare whether the efficacy differences between the models were statistically significant.P< 0.05 was considered as a statistically significant difference.Sample marker foci were randomly split on the platform and split into training and validation sets.Four sequence models,T1 WI,T2WI,DWI and SWI,were extracted to screen out the key omics characteristics.3.The HIE prognosis evaluation model was established based on T1WI-T2 WI,T1WI-SWI,T2WI-DWI,T1WI-T2WI-DWI,T1WI+T2WI+DWI+SWI multimodal sequence,and combined with NBNA score.According to the number of lesions in multimodal sequence(T1WI+T2WI+DWI+SWI),set(training set + validation set):external validation8:2 ratio,use global hyperparameter selection to extract the optimized image omics feature set,and obtain sensitivity,specificity and AUC curves.The prediction model established by multi-modal sequence model and NBNA score was used to preliminarily study the evaluation of neurological plasticity in HIE children.Results: 1.Among the 309 children in this study,the number of foci in each sequence of extracted MRI was n=1454 per unit,including training set n=1040,validation set n=172,test set n=242,and the difference between training set and test set between sex,gestational week and NBNA score.The images of MRI lesions in HIE were classified,and eight categories of four types of brain sites were classified.T1 WI,DWI and SWI sequence images in MRI image sequence were used as basic image data,and U-net deep learning model was used for image segmentation training,and then test data were selected for automatic segmentation verification.The model can accurately outline the outline of the lesion,and is relatively stable in the lesion model with different lesion properties and different lesion morphology.According to the Dice evaluation index standard evaluation,the highest similarity of the depth segmentation model is 0.832,which can reach a high coincidence degree.Among them,the Dice index of bleeding lesions,ischemia lesions,cerebral softening and brain swelling was 0.88,0.86,0.82,and0.79,respectively,suggesting that better results can be achieved compared with the deep segmentation model of bleeding lesions and ischemic lesions.The Dice indices for mass,line,spot sheet and irregular shapes were 0.87,0.82,0.79,0.70 and 0.70,respectively.It indicates that the irregular deep learning segmentation model is slightly lower than that of the mass,line and dot slice model.ACC was 0.75 in the training set,0.78 in the validation set,and 0.73 in the test set.The area under the curve(AUC)was 0.72 in the training set,0.70 in the verification set and 0.72 in the test set.2.The validation set of the neonatal HIE T1 WI sequence model that predicted NBNA scores had an AUC of 0.0.8695 and an AUC＿CI of 0.87 [0.82 – 0.92].The accuracy,sensitivity,and specificity of the prediction model in the validation set were 0.79,0.75,and 0.83,respectively.The validation set of T2 WI sequence models predicted NBNA scores with an AUC of 0.8909 and an AUC＿CI of 0.89 [0.8508-0.9309].The accuracy,sensitivity,and specificity of the prediction model in the validation set were 0.80,0.79,and 0.80,respectively.The validation set of NBNA scores for DWI sequence model predicted NBNA AUC of 0.8941 and AUC＿CI of 0.89 [0.8538-0.9343].The accuracy,sensitivity,and specificity of the prediction model in the validation set were 0.80,0.80,and 0.80,respectively.The validation set of the SWI sequence model predicted NBNA score had AUC of 0.8319 and AUC＿CI of 0.83 [0.7686-0.8951].The accuracy,sensitivity,and specificity of the prediction model in the validation set were 0.81,0.57,and 0.94,respectively.3.The neonatal HIE combination(T1WI-T2WI-DWI-SWI)sequence model predicted the ROC curves of the NBNA score,with the training set AUC 1 and AUC＿CI1[1.0-1.0].The validation set AUC is 0.9374,and the AUC＿CI is 0.94 [0.886-0.9888].The accuracy,sensitivity,and specificity of the prediction model in the training set were1,1,and 1,respectively,while the accuracy,sensitivity,and specificity in the validation set were 0.87,0.95,and 0.75,respectively.According to the first-order feature,morphological feature and texture feature,three feature types are modeled respectively.The results showed that texture features GLCM,glszm and GLDM had the best ability to predict the prognosis of HIE children in combination with(T1 + T2 + DWI + SWI),and no morphological texture features were selected.In the pairwise sequence combination and triple sequence combination models,morphological features,Laplacian Gaussian and wavelet features have better prediction ability.In the external validation set,the AUC was 0.98,and the accuracy,sensitivity and specificity were 0.83,1 and 0.71 respectively.Conclusions: 1.Based on the U-Net deep neural network model,automatic segmentation model for MRI imaging lesions in children with HIE;this model combined with 2D MRI multi-sequence image can segment relatively complete MRI lesion morphological features,laying a foundation for deep learning and radiomics research of such diseases.Based on the sample validation of the 2D MRI lesion image labeling and in the training of the U-Net deep learning model,the lesion models are relatively stable in different lesion properties and different lesion morphologies.Some lesion depth segmentation models can reach a high degree of coincidence(similarity 0.832).Among them,the segmentation effect of hemorrhage,ischemia,cerebral softening and brain swelling is better than that of mass,line,dot piece and irregular shape.2.In different types of neonatal HIE MR lesion images,T1 WI,T2WI,DWI,and SWI sequences can be displayed using machine learning techniques to extract radiomics features.Through omics feature modeling and external verification,the random forest algorithm selected based on the global hyperparameter and hyperparameter search is the optimized machine learning classification algorithm,and the prediction model established by each sequence has good stability.3.The random forest algorithm was selected based on the global hyperparameter and hyperparameter search which is the most optimized machine learning classification algorithm.In the established joint magnetic resonance multimodal prediction model,the joint(T1WI-T2WI-SWI)sequence model has the highest accuracy,and the AUC is between 0.83 and 0.92,which is a texture feature with good discrimination.In addition,the logistic regression(Logistic Regression)algorithm performs the best sequence prediction model(T1WI-T2WI-DWI-SWI),and AUC reached 0.94.The predictable HIE prognosis established with NBNA score can not only have the preliminary prediction of HIE prognosis,but also evaluate the characteristics of HIE MR imaging lesions to assess the late brain plasticity of HIE children.