Automatic Detection and Removal of Counting Chamber Grid Artifacts from Microscopy Tile Collections

Counting chamber microscopy remains the manual reference standard for quantifying biological particles in suspension. The etched reticle grid co-imaged with every acquisition tile constitutes a structured artifact that current whole-slide imaging quality control pipelines do not address. We present a preprocessing pipeline to detect and remove these artifacts via inpainting prior to phase-correlation-based alignment and stitching. To resolve the trimodal intensity distribution characteristic of brightfield counting chambers, the detection stage combines multi-level Otsu binarization with adaptive template matching for partially visible edge stubs. A Random Forest classifier trained on six geometric features filters false positives before candidates are normalized into binary masks for Large Mask inpainting (LaMa). Applied to 686 manually annotated tiles of mussel veliger larvae in a Sedgwick-Rafter chamber, the pipeline achieves a tile-level F1 of 0.9934 and a pixel-level F1 of 0.9642, outperforming a single-threshold Otsu baseline. Detection failures are confined to border stubs covering less than 2.4% of the tile dimension. The proposed configuration-driven solution adapts to varying chamber geometries via parameter updates, establishing a reproducible framework for suppressing periodic hardware artifacts in automated counting workflows.Version This record contains the preprint / author-accepted manuscript of the paper.