Segmentation-Driven Optimization For Iterative Reconstruction in Optical Projection Tomography: An Exploration

Three-dimensional reconstruction of tomograms from optical projection microscopy is confronted with several drawbacks. In this paper we employ iterative reconstruction algorithms to avoid streak artefacts in the reconstruction and explore possible ways to optimize two parameters of the algorithms, i.e., iteration number and initialization, in order to improve the reconstruction performance. As benchmarks for direct reconstruction evaluation in optical projection tomography are absent, we consider the assessment through the performance of the segmentation on the 3D reconstruction. In our explorative experiments we use the zebrafish model system which is a typical specimen for use in optical projection tomography system; and as such frequently used. In this manner data can be easily obtained from which a benchmark set can be built. For the segmentation approach we apply a two-dimensional U-net convolutional neural network because it is recognized to have a good performance in biomedical image segmentation. In order to prevent the training from getting stuck in local minima, a novel learning rate schema is proposed. This optimization achieves a lower training loss during the training process, as compared to an optimal constant learning rate. Our experiments demonstrate that the approach to the benchmarking of iterative reconstruction via results of segmentation is very useful. It contributes an important tool to the development of computational tools for optical projection tomography.