Computer Aided Diagnosis Of Alzheimer’s Disease Based On Radiomics

Alzheimer’s disease(AD)is an irreversible degenerative disease of the cranial nervous system and the fourth most common lethality disease of the elderly population in China.Because the etiology of AD has not been fully elucidated and there is no effective treatment,so early diagnosis and intervention of AD are particularly important.At present,as the clinical precursor of AD,the early diagnosis of mild cognitive impairment(MCI)has attracted more and more attention from the academic community.MCI is an intermediate stage between normal aging and dementia,in which the cognitive function of patients has a mild decline,but the ability of daily living has not received significant interference.After clinically effective intervention,MCI patients may avoid further deterioration of cognitive ability.Therefore,the early diagnosis of MCI and the prediction of MCI conversion to AD also have important practical clinical significance.Common AD,MCI diagnosis and prediction methods include biochemical indicators such as cerebrospinal fluid tau protein,Aβ42 deposition,etc.;cognitive ability based scale tests such as MMSE(Mini-mental State Examination)and CDRSB(the Clinical dementia rating scale in its sum of boxes);and using neuroimaging(eg,magnetic resonance imaging,positron emission tomography)to diagnosis.Because the biochemical indicators are difficult to obtain and they are invasive,and the scale tests will easily interfered by non-pathological subjective factors.However the neuroimaging method can avoid these problems.Therefore,in recent years,computer-aided diagnosis and prediction methods based on neuroimaging have gradually become research hotspots.At present,the common analysis methods of neuroimaging are mainly based on image feature markers(such as hippocampal volume,cortical thickness,etc.)for analysis and diagnosis,or based on machine learning and deep learning methods to construct diagnostic algorithms.However,these methods are insufficient: the image marker methods are based on the original low-order image features,and the diagnostic prediction accuracy needs to be improved.The methods based on machine learning and deep learning have a high accuracy,but its black box process will lost a lot of valuable medical information caontained in medical images.In order to solve the above problems,this study intends to adopt an emerging medical image feature extraction and analysis technology: Radiomics,which has been currently used in the study of oncology.However,the literature search results indicate that no researchers have applied the radiomic methods to the AD diagnosis related field.Therefore,the main purpose of this paper is to explore whether radimics could be applied to the AD,MCI classification and prediction.This paper firstly proposes a computer-aided diagnosis process framework for diagnosis and prediction of AD and MCI with reference to oncology radiomics.There are five steps: image preprocessing,region of interest segmentation,feature extraction,feature selection,classification and prediction.In this paper,image preprocessing,region of interest segmentation and feature selection are improved using new methods,which are different from oncology radiomics.Including registration and denoising of brain images,segmentation of regions of interest based on voxel morphology and deep learning methods.In this study,we use the pathological information and feature values to conduct medical statistical analysis,and screen out effective features with clinical analysis value.In order to verify the feasibility and clinical significance of this framework,two studies were conducted in this study.The first study was a computer-assisted classification diagnosis of Alzheimer’s disease based on 18F-FDG PET(18F-Fluorodeoxyglucose Positron Emission Tomography)images.The study included 466 samples of AD,MCI,and healthy elderly(HC).The vocal morphological method was used to determine the brain regions of interest.215 radiomic features were extracted from each sample.The stable radiomic features were determined based on Cronbach’s alpha coefficient.The Pearson correlation coefficient was used to screen out the effective features related to the cognitive scale.Finally,the support vector machine are used to diagnose the classification of AD,MCI and HC.The results showed that there were 168 stable features in brain 18F-FDG PET imaging that were not affected by random errors.The results of 500 randomized cross validation experiments shows that an average of 60 radiomic features were associated with AD cognitive ability and 35 with MCI cognitive abilities.In the final classification test,the method achieved an highest accuracy of 91.5% to distinguish between AD and HC,83.1% to MCI and HC,and 85.9% to AD and MCI.The second study was a computer-assisted predictive diagnosis from MCI to AD based on MRI images.The study included 371 stable MCI patients(188)and patients(183)with MCI converted to AD dementia.The U-Net based deep learning method was used to segment the hippocampus as the region of interest,and 520 features were extracted from each sample.The stability of radiomic features were determined,and Kaplan-Meier method was used to screen out the effective features that can reflect the conversion rate of MCI to AD dementia.Finally,the support vector machine was used to predict whether MCI patients would be converted into AD dementia.The results show that there are 414 stable features in brain MRI imaging,of which 50 effective features can reflect the risk of transformation.Eventually,a potential patient with a conversion to AD in MCI was distinguished with an accuracy of up to 93.0%.In summary,this paper proves that the radiomic methods can be used to develop computer-aided classification and predictive diagnosis methods for AD and MCI.Its accuracy is not only higher than classical image analysis methods and deep learning methods,but also can reveal the pathological information hidden in neuroimaging.This provides a new perspective for deeper follow-up research.