Convolutional Neural Networks Based Approach For Cell Segmentation And Classification

Automated cell segmentation and classification is a vital yet challenging task with significant benefits to biomedicine.Several attempts have been made to build a machine learning-based cell classifier recently using cell images obtained from a microscope.Although these studies showed promising results,such classifiers were not able to reflect the biological diversity of different types of cells in distinct microenvironments.This is thought to be closely related to cells' abnormal growth features,their ability to invade the surrounding environment,and proliferate.Therefore,segmenting and classifying different cell types based on their biological behavior using an automated technique is advantageous.Since human eyes cannot detect the features,we proposed the application of a convolutional neural network(CNN)in cell segmentation and classification based on cryomicroscopic and confocal microscopic images.The CNN was evaluated on a large number of HeLa and HUVECs cell images.The study revealed that CNN performed better in the cell segmentation and classification task than traditional approaches.The main purpose of cell segmentation and classification using machine learning is to extract the region of interest in cell images based on morphological shape information.Convolutional Neural network techniques have been widely used in cell segmentation applications because of their excellent boundary detection accuracy.In the context of medical image segmentation,weak edges and inhomogeneities remain essential issues that may limit the accuracy of any segmentation technique formulated using CNNs models.This Thesis develops new methods for the segmentation of cell images based on the CNNs models.Two different approaches are pursued.Chapter 2 proposes a Convolutional Neural Networks based U-Net architecture for HeLa Cells segmentation in the cryobiology background.This was motivated due to greater accomplishment on convolutional neural networks in the image processing field in recent decades that have outperformed the well-established protocols in the several recognition tasks,which been extensively applied in the field like semantic segmentation,image detection,and image classification.In the next chapter,we proposed a Multi-model class deep convolutional neural network for automatic HUVECs cell segmentation and classification.The model consists of two deep learning frameworks.One model is a multi-class deep CNNs for jointly learning cell segmentation and classification,which generates a binary segmentation mask of the cells and learns features for the classification process.However,the learned features by the multi-model class techniques are not sufficient for accurate cell classification.Therefore,a 3D patch covering the HUVECs cells is extracted based on the centroid of the segmentation mask and input into a 3D DenseNet model that learns more significant features for cell classification.Finally,a fully connected layer and a softmax layer are appended to combine these models' learned features for final cell classification.In this research,two classes are considered in the classification task.The input of this framework is a large image patch covering the cells.The outputs are the cells mask as well as the prediction of the cell status.