Semi-Supervised Learning With Application In MR Image Segmentation

With the gradual development of network technology and informatization, semi-supervised learning has become one of the most currently active research directions in the domain of machine learning and pattern recognition, largely motivated by new problems emerged in different real-world applications such as text information processing, natural language processing, web data mining and bioinformatics. Semi-supervised learning aims at learning from both labeled and unlabeled data collected from real world dataset where obtaining enough labels is expensive and time-consuming while unlabeled data is plentiful and inexpensive to be collected. Halfway between supervised and unsupervised learning, semi-supervised learning, which is becoming a topic of significant interest, generally falls into four categories:semi-supervised classification,semi-supervised regression, semi-supervised clustering and semi-supervised dimensional reduction. Semi-supervised classification and regression consider how to estimate a target function of classification or prediction from a limited amount of labeled examples and a large quantity of additive unlabeled examples. On contract, semi-supervised clustering and dimensional reduction employ side information of labeled examples to improve the performance of unsupervised learning.Semi-supervised learning has been widely used in many applicational fields.Image segmentation is an important issue in the field of computer vision and image procession. It refers to the process of partitioning a digital image into several different and disjointed segments each representing an object of interest. In each of these segments, image pixels should be similar with respect to local features, such as gray, color, texture and optic flow. The goal of image segmentation is to extract these regions of interest with a changed representation more easier to store, transport and analyze. Medical image segmentation refers to the analysis of medical image, such as medical radiographic image, magnetic resonance image, computed tomographic image and medical ultrasound image. The propose of medical image segmentation is to assist the programs of surgical planning, surgical navigation and image-guilded surgery, to prepare for visualization,multi-modality fusuion,multi-modality registration and interactive analysis of medical image, or to help for probing of pathology prediction and determination of morphological and structure deformations so as to diagnose diseases.Resently,one of the most active research issue is to utilize the domain knowledge in the unsupervised segmentation and information extraction, to improve the accuracy, speed and robustness of medical image segmental algorithms then ultimately make a better effect of clinical application.Cluster analysis is a basic issue in the field of data mining and pattern recognition. The propose of cluster analysis is to partitioning the whole data set based on measured similarities and dissimilarities among the classes or on the density distribution of samples and then extract useful information of interest from unlabeled data. Application of cluster analysis can be found in many fields, such as text information processing, natural language processing, web data mining and bioinformatics. Traditional clustering algorithms are always viewed as unsupervised methods for data grouping to extract information of interest from unlabeled data, while semi-supervised clustering employs limited amount of labeled data to aid the unsupervised grouping of mass unlabeled data. There are three kinds of supervised information of labeled data can be used in a semi-supervised clustering method, which are class labels,pairwise constraints and the similarity distance measurement. Based on different sources of domain knowledge, existing methods for semi-supervised clustering generally fall into three different categories:label-based methods, constraint-based methods and distance-based methods. Computer assisted medical diagnosis and medical information processing are both typical processes of semi-supervised learning, in which doctors can elect a few-typical cases to be labeled by domain knowledge and then computer program will utilize these limited amount of labeled data to aid the learning of mass unlabeled data. Recent researches in medical image segmentation intend to combine the supervised information, such as class labels, pairwise constraints and the similarity information of selected samples, with traditional segmentation methods.This paper proposes a novel semi-supervised clustering algorithm for magnetic resonance image segmentation. Pedrycz provided a semi-supervised Fuzzy C-Means algorithm (sFCM) to incorporate supervised information of labeled data as an additive part of objective function in the Fuzzy C-Means algorithm (FCM).While applied in medical image segmentation where amount of labeled samples is far fewer than that of unlabeled ones, the semi-supervised Fuzzy C-Means algorithm will turn into classical Fuzzy C-Means algorithm.In this case, the important role of labeled samples to aid unsupervised clustering will be neglected and invalid. In this paper, a novel algorithm,called Degeneracy-Improved Semi-Supervised Fuzzy C-Means algorithm (dsFCM), is proposed to fundamentally overcome the critical disadvantages of Pedrycz's sFCM algorithm, i.e., degeneracy to the classical FCM algorithm and slow convergence, particularly when applied in actual data set in which the amount of labeled points is far fewer than that of unlabeled ones.Experiments on benchmark data and real-world data show the effectiveness of the algorithm. Experimental results on UCI benchmark data and IBSR brain MR image data demonstrate that dsFCM algorithm can outperform sFCM algorithm in accuracy, speed and robustness. Moreover, it shows that dsFCM algorithm avoids the problems of slow convergence and degeneracy to classical FCM algorithm when applied to real world data clustering with exiguous labeled data, and presents its effectiveness for the application in interactive segmentation of medical images with a small amount of labeled data points given by user.