Classification of Structural and Functional Development Stage of Cardiomyocytes Using Machine Learning Techniques

The study is dedicated to the problem of classification of structural and functional development stage of cardiomyocytes derived from the induced pluripotent stem cells with application of the digital image processing methods and machine learning algorithms, in particular, neural networks. Cell regenerative therapy has become one of the most promising treatment options for patients with heart failure. But since cardiomyocytes are objects with a high level of complexity and have significant morphological variability, automatic classification is complicated by the lack of implemented methods. That's why researches in this area are a major global public health priority. The initial data set used in this study is a publicly open set of confocal microscopic images of cardiomyocytes which can be divided into five classes based on the morphological features (the structure of the transverse T-tubule). A small amount of input data leads to the need of using augmentation methods. Methods that prevent the alteration of the transverse T-tubule, which is an important parameter for correct classification of the development of cardiomyocytes, are used. Histogram equalization technique is used to enhance the contrast and dynamic range of the confocal microscopic images with the method of contrast-limited adaptive equalization. This helped to improve the local contrast of the analyzed images and highlight the main elements of the cardiomyocyte. Finally, Chan–Vese method, which belongs to the regional segmentation methods, is chosen for the image segmentation and removing artifacts and/or parts of other cells from the image. A pre-processed and augmented dataset is used for training of the convolutional neural network having an architecture with hierarchical structure and residual block usage. The model is evaluated based on the confusion matrix and the heat maps of different convolutional layers are analyzed. Images from the classes with a large number of mutual errors are also considered. Based on the analysis, several classes of structural and functional development of cardiomyocytes are combined. Final accuracy of the model for defining the cardiomyocytes maturation stage achieved 77%.