Deep learning-based image analysis with RTFormer network for measuring 2D crystal size distribution during cooling crystallization of β form L-glutamic acid

Abstract

In this paper, a deep learning-based image analysis method is proposed for in-situ measurement of two-dimensional (2D) crystal size distribution during the cooling crystallization process of β form L-glutamic acid (β-LGA). Firstly, an image quality assessment strategy is presented for in-situ snapshotted crystal images to distinguish different crystallization stages, followed by image enhancement for the snapshotted images in each stage to facilitate analysis. Then, an edge-guided network based on the RTFormer network is developed to acquire precise crystal image segmentation and boundary location, thus improving the identification accuracy on crystal image boundary and its internal body. The network performance is further enhanced by using hyperparameter optimization and a class balance strategy. Subsequently, another identification strategy is developed to distinguish agglomerated and overlapped crystal images, so as to acquire more individual crystals for statistical measurement. Finally, the 2D size of each crystal is calculated based on the major axis and maximum inscribed circle of its segmented image. Experiments on measuring the 2D size distributions of crystal populations during β-LGA crystallization process are performed to verify the accuracy and efficiency of the proposed measurement method.