Curvilinear segmentation and classification of blood vessels using random forest and deep learning
Curvilinear structures occur in many biomedical imagery, and reliable segmentation and classification is crucial for automatic image analysis. In this dissertation, we consider optimized filter bank (OFB) features with Random Forest (RF) classifier and deep learning networks to solve an important task: segmentation and pixel-wise classification for blood vessels in mice Dura Mater (DM) microscopy epifluorescence images. We consider vessel specific features which are optimized for obtaining accurate segmentations of micro-vessels under difficult imaging conditions. The segmentation is particularly challenging due to the excessive stain in the background and lack of stain within the lumen of the vasculature, different binding properties between arterioles and venules, uneven contrast, low texture content, and nonlinear binding of the fluorescence dye. By utilizing OFB features with max pooling along with a RF classifier, our segmentation method (RF OFB) is robust and avoids broken segments. We also developed and adapted new deep learning networks (VNet, SVNet, and optimized U-Net) for obtaining robust segmentations of microvasculature. As a result, we've proved that combining both RF OFB and optimized U-Net (RF OFB+U-Net) architectures achieves the highest score in DICE and overall accuracy compared with individual approaches and state of the art methods in biomedical segmentation. Our developed pipeline works on different biological imaging modalities and tissues such as dura mater, retina, and even cells. The RF can be interactively trained on regular computing hardware CPUs with small memory. As a result, the efficiency, speed, and simplicity of the infrastructure are attractive to consider our robust developed pipeline, RF OFB, as a solution for different problems.