The Study Of Color Detection On Immunohistochemistry Staining Image Based On Statistical Model

Immunohistochemistry is one of the important techniques for cancer diagnosis.Computerized digital image processing techniques can aid doctors to obtain diagnostic results objectively,quickly and accurately,and may have a significant effects on promoting automation of medical diagnosis.The color corresponding to positive staining,brown,is an important feature when analyzing immunohistochemistry staining images.In previous studies,the detecting of brown color in immunohistochemistry staining images are almost relying on color spaces.However,there is no study has shown that which color space is the best one for detecting brown color in immunohistochemistry staining images.In this paper,we study the characteristics of immunohistochemistry staining color and implemented a method with 17 existing color spaces to determine the most suitable color space.The color detection method based on statistical model has been demonstrated can effectively detect the brown color in immunohistochemistry staining images.Therefore,a set of statistical models based on each of existing color space is constructed.Different from previous studies,the method of combining the color components contained in each color space is carried out.The 17 color spaces form a total of 50 color component combinations,and 50 color detection statistical models are trained on the same data set.Experimental results have shown that the most suitable combination is rg-by.Compared with the classical and the latest color detection methods,statistical model with rg-by has obtained the highest detection accuracy.The distribution of brown in the immunohistochemical staining image dataset produced by different laboratories is quite different in the color space.Previously studied methods for brown color detection vary widely in the accuracy of cross-dataset detection.Although the statistical model based on the combination of rg-by color components has a high detection accuracy on the current data setthe detection accuracy is still significantly reduced with application to a new data set.In order to solve this problem,this paper proposes a new method to develop a hybrid color space by using multiple existing color spaces.Due to the high dimensionality of the hybrid color space created by using multiple color spaces,it is not suitable for practical applications.In this paper,we firstly use the four color components that are most sensitive to brown to establish a hybrid color space,then the principal component analysis method has been used to reduce the dimension of this hybrid color space,finally the first two columns are regarded to be a new hybrid color space.The statistical model established based on the new hybrid color space has been examined on three pathological image datasets obtained from the human pathological sections of esophageal,colorectal and liver cirrhosis biopsy.The experimental results have shown that,as comparing with all the most commonly used brown color detection methods,the statistcal model based on two-channel hybrid color space has obtained higher detection accuracy that equal to or slightly higher than other brown detection methods,and lower detection accuracy variations on three public datasets.Finally,the positive stainging color detection method proposed in this paper has been developed into a brown detection tool.The tool has good detection accuracy when detecting brown color,and the detection accuracy is stable when detecting across datasets.