A Deep-Learning Approach to Marble-Burying Quantification: Image Segmentation of Marbles and Bedding

This paper presents and evaluates three automated tools for semantically segmenting images from marble-burying experiments. The marble-burying animal model is widely used as an indication of anxiety or obsessive compulsive behavior in rodents. In general, the tendency for caged rodents to bury objects in their bedding is seen as anxiety related, and several methods have been proposed to measure the degree of this burying behavior. Unfortunately, most of these methods are coarse or require either subjective interpretation or onerous manual procedures. Digital imaging can provide pre- and post-experiment burying states as well as a platform for a standardized and streamlined quantification of the marble-burying test. While continuous imaging streams might provide more information and temporal analysis, such datasets are rare, require expert annotation, and can be prohibitively large. The authors propose that single-image semantic, pixel-wise segmentation of marble and bedding pixels are key components that can enable effective quantification. For example, the ratio of marble to bedding pixels can provide greater granularity in assessing marble-burying behavior. In this work, a classical image segmentation approach, a single-class U-Net and a multi-class U-Net were comparatively evaluated via standard segmentation metrics. Results show that the deep-learning methods demonstrate greater segmentation performance than the traditional method. Timing-performance trade off considerations between single- and multi-class methods are also explored.