Background: Mold identification in clinical diagnostics is traditionally labor intensive and is dependent on expert interpretation. MoldVision is a deep learning approach that uses smartphone images ofmold cultures to automate identification.

Methods: We analyzed 161 clinical isolates across four common mold genera. Penicillium spp., Aspergillus spp. (with A. flavus and A. fumigatus), Fusarium spp., and Cladosporium spp. Daily images were captured from the top and bottom of culture plates over five days using a standardized smartphone setup, generating over 4,000 images. We trained three variations of VGG16 convolutional neural networks (CNN) and benchmarked the best-performing model (VGG16 with dual classification heads) against LightGBM models trained on pre-extracted features and human expert assessments at various time points.

Results: The best performing VGG16 model achieved a mean (SD) ROC-AUC of 92.7% ± 1.8% and sensitivity of 68.7% ±2.6% across all species. Here, the performance in identifying Cladosporium spp. was best (ROC-AUC 99.9% ± 0.1%, 5-fold cross-validation mean and SD ). Regarding the evaluations over time, early stage classification (days 1-2) was challenging (F1-score 38.8% ± 3.5% across all species but improved significantly on day 3-5 (F1 92.1% ± 2.8% across all species). Compared to experts, MoldVision consistently showed superior performance, particularly in mature cultures, detecting subtle morphological features earlier and more accurately.

Conclusions: Our results demonstrate that CNNs integrated with low-cost smartphone imaging can reliably classify mold species in routine diagnostics, outperforming human experts in many cases. This approach offers a practical and scalable solution for laboratories lacking specialized mycology expertise, especially in resource-limited settings.