Abstract
The presence of dendritic cells (DCs) introduces significant biases in the assessments of corneal nerve morphology, primarily leading to false-positive results in dry eye disease (DED), especially chronic FS-LASIK related DED. Their exclusion improves the precision of nerve measurements, which may enhance the clinical evaluation of corneal nerve morphology. These findings highlight the importance of precise segmentation techniques to minimize DC-related interference in clinical practice.
Introduction
In vivo confocal microscopy (IVCM) serves as the gold standard for cellular-level imaging of corneal nerves. However, a significant diagnostic challenge exists: dendritic cells (DCs)—immune cells that proliferate during inflammation—often reside near these nerves. Because both structures share similar refractive properties and branching shapes, standard automated software frequently misidentifies DCs as nerve fibers.
This "masquerading" effect is particularly pronounced in patients with chronic Dry Eye Disease (DED) following Femtosecond Laser-Assisted Laser in Situ Keratomileusis (FS-LASIK), where inflammation significantly increases DC density. This study quantifies the resulting bias and evaluates how excluding DCs alters the understanding of nerve morphology in clinical contexts.
Methods and Materials
The researchers conducted a cross-sectional study involving 61 participants, categorized into healthy controls (n=24), patients with DED without LASIK (n=20), and patients with chronic DED after FS-LASIK (n=17). High-resolution images of the central cornea were captured using a Heidelberg Retina Tomograph 3.
To differentiate between nerves and immune cells, a novel end-to-end encoder-decoder AI architecture was developed. The model utilizes a ResNet50 backbone with two specialized branches:
- Nerve Branch: Employs a Multi-Scale Fusion Module (MSFM) to capture fibers of varying sizes.
- DC Branch: Implements an Uncertainty Boundary Guided Module (UBGM) to handle the "fuzzy" boundaries of DCs, allowing for precise segmentation and exclusion of these cells from nerve metrics.
Key Insights: The "False Positive" Effect
The study revealed that including DCs in morphological analysis creates an illusion of higher nerve density and complexity. This bias is most severe in post-LASIK DED patients due to their high DC density. In this group, the "noise" from DCs was sufficient to create statistically significant—but ultimately false—correlations between immune cell presence and nerve structure.
Furthermore, in the DED-only group, including DCs led to false-positive correlations between nerve morphology and clinical characteristics like corneal sensitivity. This suggests that previous research may have misattributed clinical symptoms to nerve structural changes when they were actually observing the footprint of an immune response.
Key Results
Excluding DCs led to a significant reduction in all measured nerve metrics across all study groups. Specifically, Corneal Nerve Fiber Density (CNFD) decreased from 4.69 ± 1.54 /mm² to 4.29 ± 1.69 /mm² ($p < 0.001$).
Other metrics, such as fractal dimension (Boxdim) and tortuosity, also saw significant reductions. While the post-LASIK DED group consistently showed lower nerve density and higher tortuosity than other groups, the true magnitude of these differences was only accurately revealed after DC exclusion. Notably, the post-LASIK group exhibited significantly higher DC density than the non-LASIK DED group ($p < 0.01$), highlighting the long-term inflammatory impact of the surgical procedure.
Conclusion and Future Directions
This research demonstrates that precise AI-driven segmentation is essential for accurate corneal nerve assessment. Traditional methods that fail to exclude DCs risk overestimating nerve density and misinterpreting the relationship between nerve structure and clinical symptoms.
While limited by the IVCM's field of view and a focus on structural rather than functional signaling, these findings advocate for the integration of advanced segmentation techniques into standard clinical software. Future work should explore these biases across a broader range of ocular surface diseases to ensure diagnostic accuracy in neuroimmune evaluations.
Summary updated