Research And Implementation Of Classification Algorithms For HIV Brain Images

Human Immunodeficiency Virus(HIV),also known as AIDS virus,is a retrovirus that causes defects in the human immune system.Within one week of HIV infection,even after receiving antiretroviral therapy,at least 50%of infected individuals will still experience sensory and motor dysfunction and neurocognitive impairment,known as HIV-Associated Neurocognitive Disorders(HAND).Patients with HAND may experience memory loss and delayed reactions in the late stages of infection.Since HIV invades the brain and causes irreversible damage to the nerve cells in the brain,early diagnosis is important for the treatment outcome and the life quality of patients.Therefore,how to accurately diagnose the early stages of HAND is an important research direction.At present,the detection of HAND still depends on doctors’subjective judgment,which is called the neuropsychological scale method.Therefore,there is an urgent need to establish a more objective diagnostic method,such as early diagnosis through objective information such as medical images and blood indicators.Meanwhile,due to the scarcity of medical professionals and the rapid development in the field of computer vision,computer-aided diagnosis using medical images is a research direction of great significance.In summary,a study on brain medical image classification algorithms was carried out in this thesis to address the problem of computer-aided diagnosis in the early stages of HAND.One of the specific technical objectives is to design a model for classifying medical brain images of the HAND patients and non-diseased subjects.The core problem encountered in the research process is that the small sample size of the HAND dataset makes it difficult to train a classification model with good performance using traditional deep learning methods.Few-sample is also a common problem in the field of medical image assisted diagnosis.To address the problem of small sample size in the HAND dataset,this thesis is based on the idea of meta-learning,in which various classification tasks are used as training sets for learning,and then the classification ability learned from the training tasks is generalized to the target classification tasks,thus reducing the dependence on the number of training samples.However,the optimization obtained by tasks of different complexity in traditional meta-learning methods is uneven,which leads to the parameters of complex classification tasks often cannot be fully optimized.To address this problem,this thesis proposes a meta-learning model that can dynamically adjust the task weights to give more optimization to complex tasks in the meta-learning process,so as to improve the classification effect of the model.The experimental results show that the classification model based on meta-learning only needs a few training samples to obtain better classification performance than the traditional model,but this advantage only stays in the case of few samples.Therefore,this thesis turns to a data augmentation approach to solve the problem of insufficient number of samples in the HAND dataset.A multiscale fusion module is also designed to sample the brain medical images with multiscale information,which solves the problem of unclear HAND lesion regions.Finally,it is experimentally shown that the multiscale 3DCNN classification model achieves the optimal classification performance compared with the previously proposed classification models.