Transfer Learning Algorithm For HIV Brain Image Classification

With the development of information technology,medical imaging has become an important carrier for obtaining clinical information.Medical images are often used in clinical diagnosis due to their high resolution,richer information and advantages in improving efficiency.The improvement of medical standards has placed higher requirements on medical image analysis technology,so traditional methods of clinical analysis cannot meet the needs for accurate detection and rapid diagnosis of lesion features Therefore,scholars began to try to apply computer-aided diagnosis methods to medical image analysis of diseases.AIDS virus is a virus that destroys the immune function of the human body.Due to the insufficiency of clinical screening and treatment methods,it often leads to disability or death of HIV-infected patients.The symptom of brain structural change and cognitive impairment in HIV-infected patients are called HIV-associated neurocognitive disorders(HAND).In order to solve the problem of early diagnosis of HIV-infected persons,this article conducts research on HAND brain imaging.The main contents are as follows:In response to the problem of overfitting caused by the small amount of brain image data of HIV-infected persons,a brain image classification model based on transfer learning and attention mechanism is proposed.The features learned from the 3D images of various diseases in the medical field are used in the HAND brain image classification experiment.In order to measure the influence degree of features on the HAND classification results,this paper introduces an attention mechanism.The features are further selected and the convergence is accelerated when giving different weights to each channel.In addition,since the performance of patients’cognitive and motor abilities is related to changes in brain structure,this paper will use the features extracted from HAND brain images to predict the patient’s cognitive abilities.Under the combination of HAND brain image and normal brain image classification and the subject’s cognitive ability prediction,the model has good performance.The HAND brain image classification can reach an accuracy of 88.9%.In the brain classification model based on the depth transfer and attention mechanism,the model is obtained by monomodal brain image training.However,since only a single modality is used for feature extraction,it is easy to ignore the characteristics of other modality.In order to classify HAND brain images more comprehensively,this paper uses two modalities of high-resolution structural MRI and resting brain functional MRI.Experiment with data.First,select Alzheimer’s disease brain image data similar to HAND symptoms for data augmentation and model training,and then migrate the pre-training model parameters to the HAND multi-modal brain image classification task.By combining multi-modal information,the model can achieve an accuracy of 90.3%.