Stem cell detection based on Convolutional Neural Network via third harmonic generation microscopy images

Stem cell plays an important role in repairing destroyed tissues and keeping human healthy every day; and thus stem cell observation and detection are principle procedures before being analyzed by physicians. In this paper, we proposed a criterion of Computer-Aided Diagnosis (CAD) system to detect stem cells in the stratum basale based on cell segmentation algorithm and intrinsic characteristics of cells, which can provide consistent and accurate results for assisting the assessment of diagnosis. In addition, we utilize Convolutional Neural Networks (CNNs) to recognize basal cells and stem cells since CNN has excellent performance on processing abundant data. Actually, the procedure of acquiring biomedical images is too complicated to collect, hence hand-crafted initialization is adopted to overcome the issue of the lack of training data according to prior knowledge or the suggestion from medical doctors. The experimental results show that the accuracy of hand-crafted initialization is higher than random distribution kernels and the convergence time is shorter also since a better initial condition may lead to better results in optimization theory.