From annotation to analysis: a deep-learning pipeline for optical coherence tomography (OCT)-based measurements of biofilm morphology.

Biofilms represent the predominant mode of bacterial life at solid-liquid interfaces, and understanding their composition, structure, and dynamics is critical to addressing key challenges across medical, environmental, and engineering applications. This study presents a deep learning-based framework for rapid morphological characterisation of biofilms using optical coherence tomography (OCT) imaging and an automated image processing pipeline. Images were used to train two state-of-the-art segmentation models: YOLOv8 and SegFormer. Both models delivered impressive results in delineating biofilm structures; YOLOv8 achieved 0.99 for accuracy and an intersection over union (IoU) of 0.9, while SegFormer scored 0.97 and 0.87, respectively. Model robustness was assessed across eight challenging biofilm conditions, with YOLOv8 showing superior performance in discriminating thin and non-growing biofilms, and SegFormer's superiority with stable morphologies. Additionally, we developed a framework to extract key morphological characteristics from the segmented images, including thickness, roughness and density distribution. The model-derived measurements showed strong agreement with manually generated ground truth data, confirming the reliability of the automated pipeline. Furthermore, an experiment involving four taxonomically distinct multi-species biofilms demonstrated the utility of the approach for discriminating biofilms based on their morphology. The software, containing both segmentation models, is openly available to the community and provides a foundation for future high-throughput studies examining biofilm responses to taxonomic or environmental variation.