Computer-Aided Analysis of Scanning Electron Microscopy Images with Illumination Inhomogeneity and Touching/Crossing Cells

Abstract

Clostridioides difficile infection (CDI) is a significant cause of death and morbidity due to infec- tious gastroenteritis in the USA. Treatments for CDI are being developed and comparison of the treatments is of paramount importance. Conventional microbiology methods investigate the effec- tiveness of treatments on the macro-level, and a phenotypic investigation has not been performed. Phenotypic features (e.g., length, shape deformation) of CDI cells in scanning electron microscopy (SEM) images indicate critical information about cell health in CDI research studies. However, analysis of SEM images is challenging due to the following challenges: (1) inhomogeneous illumi- nation, which causes shadows on the cells and bright areas around the cells, and (2) the presence of touching and crossing cells. Therefore, there is an urgent critical need to develop methods for the segmentation of the CDI cells to extract phenotypic information. This work presents a deep learning pipeline to provide instant-level segmentation of CDI cells in scanning electron microscopy images. The components are: (i) an adversarial region proposal network to compute cell candidate bounding boxes, and (ii) an instance-level segmentation network extracting features from bound- ing boxes, and computing the segmentation masks of isolated, touching, and crossing cells. The pipeline provides a computational tool for analysis of scanning electron microscopy images which is critical to compare the efficacy of CDI treatments. Finally, the performance is evaluated and compared to the state-of-the-art in instant-level object segmentation. The results indicate that the proposed computational tool out-performs the state-of-the-art method Mask-RCNN in detection (mean average precision) and segmentation (dice score) of CDI cells in SEM images.