Bacteria Detection in Optical Endomicroscopy Images Using Synthetic Images

Pneumonia, a lung-related illness often resulting from bacterial infection, requires quick and accurate identification, especially in intensive care situations. Optical endomicroscopy (OEM) offers a solution by providing real-time acquisition of in vivo and in situ optical biopsies, enhancing the speed of bacterial identification. However, the sheer volume of images produced by OEM for real-time analysis poses a significant challenge, potentially delaying critical treatments. Prior approaches to bacteria detection in OEM imagery have relied on unsupervised models. These models are hindered by either the need for manual threshold setting or high computational demands, making real-time analysis unfeasible. To address these challenges, supervised learning methods can be considered, as they have shown superior performance and efficiency in various medical applications. However, supervised learning approaches heavily depend on the availability of vast quantities of accurately labeled data, which is scarce for OEM images. To this end, in this paper we introduce a novel approach to generate synthetic bacteria within OEM image sequences, enabling the use of deep learning techniques. We developed two models to simulate bacterial movement and embedded these models into real, bacteria-free background images. To assess the efficacy of synthetic image sequences, we employed a 3D U-Net for training. The results revealed that the 3D U-Net, when trained on these synthetic sequences, exhibited a 3.86% enhancement in correlation with real annotations over state-of-the-art bacteria detection models.