A kernel-based adaptive fuzzy c-means algorithm for M-FISH image segmentation

Multicolour fluorescence in-situ hybridization (M-FISH) images can be used to detect chromosomal abnormalities, which subsequently can be used for the diagnosis of certain cancers and genetic diseases. However, there is currently no automated system able to achieve a sufficient accuracy for clinical purposes. The accuracy of segmentation and classification of pixels within these images is highly important but is often marred by local intensity inhomogeneities and noise. A kernel based local adaptive fuzzy c-means (KAFCM) as well as a probabilistic defuzzification classifier were developed to improve the segmentation and classification of chromosomes. This is achieved by using a gain field over a local window for each pixel to compensate for the intensity inhomogeneities caused during the imaging process and by the physical chromosome preparation itself. The algorithm was tested on a publicly available dataset and was compared with the traditional fuzzy clustering algorithm and to the reported results of another adaptive gain field algorithm. Based on these experiments the proposed system showed comparative results. Furthermore, the classification results for both the proposed method and standard FCM defuzzification were compared, and the proposed classification method demonstrated an overall increased performance.