Functional and Evolutionary Morphology Research Lab

This lab focuses on quantifying data from various imaging sources, with a great deal originating from high-resolution CT scans of archaeologically recovered material. High-quality image segmentation (classifying types or groups of material in an image) is perhaps the most critical step in these types of analyses. The rapid rise of machine learning has allowed researchers to move away from time-consuming manual segmentation, which often introduces a human bias that will affect the results. Even with the advantage of these new methods, training deep learning networks for image segmentation takes a lot of training images. The aims of deep learning research in this lab are to provide researchers with a freely available set of tools that will provide a low/no-code method to train custom segmentation networks with a minimal set of training data, which will, in turn, decrease the reliance on manual segmentation and increase their ability to produce consistent segmentations. This method leverages expert knowledge (i.e. domain-specific) of difficult archaeological and paleontological material into the training of an ensemble U-net network. This approach has been shown to more closely match segmentations conducted by an expert (~94% overlap) when compared to other automated segmentation (~89%) and other deep-learning approaches (~92%).