Use of apriori information for anatomical segmentation of medical images

In this thesis, we investigated how in medical image segmentation, apriori information about objects’ shape knowledge can be used to guide the lungs segmentation from chest X-ray images. We propose a deep convolutional neural network for lung segmentation, where shape information is represented by a statistical representation, based on principal components analysis of lung masks computed using the training dataset and described by their Signed Distance Functions. In this regard we provide a nonlinear extension, considering the expediency of KernelPCA and the related kernel functions. As a result, the proposed network learns to predict shapes instead of learning pixel-wise classification. The segmentation method named Shape Predictor Network (SP-Net) was applied to chest X-ray images of Covid-19 positive patients where shape could be of utmost importance. Results showed that SP-Net could constrain the predicted shape to resemble a pulmonary-like structure in those cases where the pixel-wise segmentation methods failed in the presence of a significant image artifact. Finally, we compared SP-Net with state-of-the-art methods. U-net, a convolutional neural network particularly effective for medical image segmentation.