An Adaptive Semi-supervised Deep Clustering And Its Application To Identifying Biotypes Of Psychiatric Disorder

In recent years,semi supervised deep clustering has been applied in many fields and achieved very good results.Compared with traditional clustering methods,deep learning can automatically extract features,while semi-supervised clustering can use existing features of data to learn to improve clustering performance.The main work of this dissertation is to use semi-supervised deep clustering to cluster a variety of image data and apply it to the identification of psychiatric disorders biotypes.In the field of image processing,many researches use deep learning and semisupervised clustering to explore and distinguish different kinds of image data.In response to the problem that some existing semi-supervised clustering researches fail to take full advantage of the information in unlabeled data,the first work of this dissertation proposes an adaptive semi-supervised deep clustering method,which makes full use of a small amount of labeled data and extracts information adaptively from unlabeled data with high confidence,so as to facilitate the subsequent semi-supervised clustering.In order to prove the effectiveness of our method,we apply this method to cluster on Fashion MNIST,MNIST and USPS datasets.We also make a detailed comparison with K-means clustering,constrained seed K-means clustering and unsupervised deep clustering.The results show that the adaptive semi-supervised deep clustering has the best clustering effect.Our work shows that adaptive semi-supervised deep clustering can effectively cluster different datasets.In the field of brain science,the symptoms of different types of psychiatric disorders overlap,which makes diagnosis challenging.There is already evidence that exploring new biotypes may help to solve this problem.Clustering has been used to identify biological types using brain brain magnetic resonance imaging(MRI)data,but deep learning has not been well studied for this purpose.The second work of this dissertation is to apply adaptive semi-supervised deep clustering to the biological recognition of psychiatric disorders.We study the biotypes of 137 patients with autism spectrum disorder(ASD)and 137 patients with schizophrenia(SZ),and prove that the modified method can well identify the biotypes of different patients.