Objectives. We developed and tested multiple computer-vision image classifiers for their ability to identify a large set of opportunistic and pathogenic molds by their colony appearance. Aim of the study was to create a comprehensive global benchmark towards computer vision driven diagnostics of pathogenic microbes. If successfully implemented, high resource clinical settings could greatly benefit from this adjunct technique to supplement molecular sequencing and mass spectrometry diagnostic methods, while under low resource conditions in rural and remote geographies, it could provide rapid approximate pathogen identification. Methods. We selected 114 clinically relevant fungi represented by 123 strains obtained from the images implemented within the ‘Atlas of Clinical Fungi’, to serve as core dataset. The image classifiers were designed with a rigorous testing and evaluation strategy, at a yet unprecedented level of detail. We designed the framework within the TensorFlow environment, testing multiple transfer-learning approaches, as well hybrid architectures comprising features of convolutional neural networks (CNN) and advanced vision transformers (ViT). Results. We achieved a global identification accuracy of > 88 % for the validation partition with our best model (Test accu. 87 %, Train. accu. 96 %). Simulations indicated that extended training time would lead to further improvement of accuracy, particularly with greater data richness. Our results highlight complex de-black-boxing approaches in interpreting image classification, useful for validating computer-vision driven microbial identification. Discussion. Besides quantitative limitations of representative strains per tested species, our approach reflects a significant scientific novelty, providing value to the technically simple culturing technique. The method was tested on filamentous molds and a small subset of common bacteria as a control set. The methodology can be universally applied to clinical microbes, rendering the technique attractive for rapid diagnostics and verification of identities from experimental approaches.
Stielow et al. (Sat,) studied this question.