Automated 3D Reconstruction of Mycorrhizal Fungal Structures from Bright-Field Microscopy Using Oriented Bounding Boxes and Focus Modeling

Abstract Background Fungal networks play a fundamental role in soil ecology and plant symbiosis, yet their three-dimensional characterization remains challenging due to their complex filamentous morphology and the limitations of conventional microscopy image analysis. This work presents a data-driven framework for three-dimensional reconstruction of fungal structures from bright-field microscopy, combining YOLO-based oriented bounding box (OBB) detection with continuous focus modeling for depth estimation. Addressing a problem explored by relatively few existing microscopy-based reconstruction approaches, the framework was developed using images of the arbuscular mycorrhizal fungus Rhizophagus irregularis , with additional qualitative applications to other biological structures exhibiting similar filamentous morphologies. Results The OBB detector achieved an mAP50 of 0.60 ± 0.03, with instance-level and spatial-coverage F1-scores of 0.67 ± 0.02 and 0.78 ± 0.02, respectively. To account for the continuous nature of filamentous structures, a composite score combining both evaluation criteria was proposed, yielding S = 0.72 ± 0.01. Focus was modeled using complementary directional, spectral, and intensity descriptors integrated through PCA, achieving an F1-score of 0.90 on the ×10 validation dataset. The complete framework was applied to a 25-hour time-lapse experiment, enabling repeated three-dimensional reconstruction of the evolving fungal structure. Quantitative spatial evaluation yielded a global pixel-wise recall of 0.93. Furthermore, 75.75% of the focused OBBs overlapped the manual segmentation by at least 20% of their area within an axial tolerance of ± 35 µm, as measured by the axial-tolerant overlap rate (ATOR), substantially exceeding spatially randomized controls. Conclusions These results support the feasibility of combining orientation-aware detection and focus modeling for automated three-dimensional characterization of fungal structures, providing a foundation for future investigations of fungal network architecture and growth dynamics.