Research Of Software System For Chromosome Image Analysis

Chromosome image analysis is one of the essential tasks in cytogenetics, especially in genetic syndrome diagnosis. In this thesis, an automatic analysis and recognition system is introduced for human G-banded chromosome images. Some important technologies and core algorithms used in the system are discussed in detail, including the algorithms of image pre-processing, automatic segmentation of touching and overlapping chromosomes, extraction and rotation of medial axis, projection of band patterns, feature measurement, selection and classification.Automatic segmentation of touching and overlapping chromosomes is one of the most important parts of the system. According to the different status of touching or overlapping, we present a segmentation algorithm combining the curvatures of contours and the media axis distributions of chromosomes. This algorithm can reduce the times of interaction and boost the automatization.Extraction of medial axis of chromosomes is also one of the most important parts of the system. We propose a method for medial axis extraction combining the algorithms of thinning, branch removing, curve fitting and optimized curve extending. By the method, satisfactory medial axis can be extracted. The method is the base of contour projection, gray projection and gray distribution.Finding the centromeres and band pattern features is another important part of the system. We propose an algorithm for centromere finding according to the contour projection and the gray distribution. We also propose an algorithm for band patternfinding using gray projection.Automatic chromosome pairing is also an important part of the system. We propose a method of automatic pairing by classification using Bayesian classifier with 24 features. The 24 features are extracted or selected from the length, numbers of band patterns, contour projection, gray distribution, gray projection, energy of gradient projection and the six FFT coefficients of chromosomes. The automatic pairing by the method is satisfactory.There are several innovations in this paper: (1) An algorithm is proposed for curve fitting and extending of thinned medial axis. With the algorithm, the accuracy of medial axis detection is increased. (2) An algorithm is proposed for centromere detection applying the combination of the width of contour projection and the single peak of gray distribution. With the algorithm, the accuracy of contromere detection is increased. (3) A method of classifier design is proposed by combination of the spatial and frequency features. With the method, much better classification is achieved. (4) An algorithm is proposed, in which cells are treated as classifiers and the numbers of chromosomes in each class are taken account. This algorithm can reduce remarkably the error rate of classification.