COFI - Coarse-semantic to fine-instance unsupervised mitochondria segmentation in EM

nstance segmentation is crucial for insightful analysis in the increasing use of large-scale electron microscopy (EM) to gain a better understanding of disease causes or progression. Instance segmentation is a more granular version of semantic segmentation, as it identifies and distinguishes individual object instances, whereas semantic segmentation only identifies object classes. In this study, we introduce a two-stage unsupervised approach called COFI, which stands for Coarse-Semantic to Fine-Instance segmentation, for the application of mitochondria segmentation in large-scale 2D EM images. In its first stage, it produces a rough region mask by clustering image patches and prompting a user to select the regions of interest. This is followed by a boundary delineation method based on the brain-inspired COSFIRE filter which is augmented by an inhibition component that makes it robust to image texture and noise. The effectiveness of the proposed COFI approach is evaluated on an EM dataset of the heart muscle of a mouse tissue, which consisted of four tiles of 16384×16384 pixels, containing a total of 2287 instances of mitochondria among other subcellular structures. It consistently achieved panoptic quality measures that are substantially superior to competing supervised methodologies. Besides its elevated effectiveness, the proposed COFI approach is conceptually simple and sufficiently versatile as the structure of interest is not intrinsic to the method.