Researcher(s)
- John Bean, Biomedical Engineering, University of Delaware
Faculty Mentor(s)
- Logan Hallee, Graduate Student, University of Delaware
- Jason Gleghorn, Biomedical Engineering, University of Delaware
Abstract
Biomedical image segmentation is a fundamental aspect of cell biology; however, hand segmenting images is time and resource intensive. Convolutional neural networks (CNNs) provide an excellent solution for segmentation tasks given that they have access to enough relevant training data. Although, as stated previously, generating sufficiently large sets of hand segmented images is not always feasible. Furthermore, microscopy images tend to be too large to feed through convolutional layers on conventional computers. To solve both problems we implemented a “Windowed” approach to segmentation and reconstruction of microscopy images. By splitting training images into a grid of smaller panels we were able to sufficiently train a CNN on only a handful of segmented images. Not only did this approach reduce the resource commitment to create a training set, but it also lightened the computational intensity of training and practical use. We have extended this methodology to create a user-friendly pipeline to apply CNN architectures to segmentation and classification tasks. Our pipeline presents an accessible solution that produces high-quality results much faster than conventional segmentation methods. The user is only responsible for organizing their images into the file-tree structure outlined in our manual and configuring relevant settings which have been paired with our richly detailed instructions. Our pipeline then windows their images, trains and validates a CNN model, reconstructs the desired images in 2D. For spatial Z-stack imaging data, our pipeline also compiles the 2D images into a fully segmented point cloud for each class through advanced 3D interpolation. These tools enable 3D anatomical models and will have significant impacts across numerous applications including tissue morphometry, morphodynamics, and spatial pharmacology. Furthermore, our pipeline makes CNNs accessible to areas where datasets are limited, allowing more users to participate in similar studies.