Low Rank And Sparse Representation Based Mitotic Cells Detection On Breast Cancer Histopathology Images

Mitosis count in histopathology image has been one of the important indicators for breast cancer grading. It is usually time consuming and tedious to manually detect and count the number of mitosis and the results may vary largely among pathologists. Even for the same pathologist, the results may different at different time or under different working condition. So it is eager to have a computer aided method for automatic mitosis detection in the clinical practice. Essentially, the problem of mitotic detection involves irregular shape object classification. It is a very difficult task. Complicated features need to be extracted for the existing pattern recognition based methods in the literature. It takes so much time to train the classification models. To address this problem, this paper presents a low rank representation based computer aided method for mitosis detection. Mitosis cells can be considered as the sparse section while the non-mitosis can be seen as the low rank section. We can extract the mitotic cells directly from image sequence with gray features in different spectral bands. Experimental results on ICPR2012contest mitosis image set show that our method can achieve F-measure0.59and Recall0.56, respectively. The evaluation results are better than the traditional pattern recognition based methods with gray features. When we use higher level texture features to train classification models for pattern recognition based methods, our proposed method also get the best F-measure record.