A Registration and Deep Learning Approach to Automated Landmark Detection for Geometric Morphometrics

1. Geometric morphometrics is the statistical analysis of landmark-based shape variation and its covariation with other variables. Over the past two decades, the gold standard of landmark data acquisition has been manual detection by a single observer. This method has proven accurate and reliable in small-scale investigations. However, big data approaches are increasingly common in biology and morphometrics among other fields. Large-scale analyses of morphological variation require automatic, standardized, and integrative landmark data collection. Image registration, or the spatial alignment of images, is a fundamental technique in automatic image analysis that is well-poised for such purposes. Unfortunately, in the few studies that have explored the utility of registration-based landmarks for geometric morphometrics, relatively high or catastrophic labelling errors around anatomical extrema are common. These errors can result in misleading representations of the mean shape, an underestimation of biological signal, and altered variance-covariance patterns. 2. We combine image registration with a deep and domain-specific neural network to automate and optimize anatomical landmark detection for geometric morphometrics. Using micro-computed tomography images of genetically and morphologically diverse mouse skulls, we test our landmarking approach under a variety of registration conditions, including different non-linear deformation frameworks (small vs. large) and atlas strategies (single vs. multi). 3. Compared to landmarks derived from conventional image registration workflows, our optimized landmark data show significant reductions in detection error at problematic locations (up to 0.63 mm), a 36.4% reduction in average landmark coordinate error, and up to a 45.1% reduction in total landmark distribution error. We achieve significant improvements in estimates of the sample mean shape, sample-wide distance relationships, and variance-covariance structure. 4. Our approach only requires intensity information and landmark data from volumetric images. It is thus generalizable to other volumetric imaging modalities, anatomy, and landmark configurations. For biological imaging datasets and morphometric research questions, our method can eliminate the time and subjectivity of manual landmark detection whilst retaining the biological integrity of these expert annotations.