# Neural network-based image analysis of co-localized microorganisms and human cells on implant materials

## TL;DR

This paper introduces a new image analysis workflow using neural networks to study interactions between microorganisms and human cells on dental implants.

## Contribution

The study introduces a novel workflow using two custom Cellpose models for analyzing co-localized microorganisms and human cells on implant materials.

## Key findings

- A first Cellpose model effectively analyzed individual bacteria in 3D implant-tissue-oral biofilm co-culture images.
- A second model trained for microcolony recognition enabled automatic segmentation of in situ study images.
- The workflow's accuracy can be improved with more training data and better image quality.

## Abstract

Dental implant-associated infections increase the risk of implant failure, presenting significant challenges in modern dentistry. The host-microbe interaction plays a crucial role in the development of implant-associated infections. To gain a deeper understanding of the underlying mechanisms, numerous studies have been conducted using in vitro co-culture models of bacteria and human cells or in situ samples. Due to the complexity of the images generated throughout these studies, however, the analysis by means of classical image processing techniques is challenging. This study proposes a workflow—based on two custom Cellpose models—that, for the first time, allows the analysis of microbial surface coverage in microscopy images of fluorescent-stained and co-localized microorganisms and human cells with substantial background signals. The first Cellpose model demonstrated its efficacy in the analysis of individual bacteria within images derived from an 3D implant-tissue-oral biofilm in vitro co-culture model. In combination with the second custom model, which was trained to recognize microcolonies, images obtained from an in situ study could also be automatically segmented. The model’s segmentation accuracy could be further enhanced by acquiring additional training images and improving image quality, making the proposed workflow now valuable for a range of dental implant-related and other co-culture images.

The online version contains supplementary material available at 10.1038/s41598-025-05484-1.