This study introduces an automatic deep-learning-based approach to diabetic retinopathy (DR) severity assessment by integrating two modalities: Ultra-Widefield Color Fundus Photography (UWF-CFP) from the CLARUS 500 device (Carl Zeiss Meditec Inc., Dublin, CA, USA) and a comprehensive set of clinical data from the EVIRED project. We propose a framework that combines the information from 2D UWF-CFP images and a set of 76 tabular features, including demographic, biochemical, and clinical parameters, to enhance the classification accuracy of DR stages. Our model uses advanced machine learning techniques to address the complexities of synthesizing heterogeneous data types, providing a holistic view of patient health status. Results indicate that this fusion outperforms traditional methods that rely solely on imaging or clinical data, suggesting a robust model which can provide practitioners with a supportive second opinion on DR severity, particularly useful in screening workflows. We measured a multiclass accuracy of 63.4% and kappa of 0.807 for our fusion model which is 2.1% higher in accuracy and 0.022 higher in kappa compared to the image unimodal classifier. Several interpretation methods are used to provide practitioners with an inside view of the workings of classification methods and allow them to discover the most important clinical features.
Vision foundation models for retinal imaging are large and computationally demanding, limiting point-of-care deployment. We ask how much of such a model's downstream utility survives compression into a smaller network. We propose a topology-aware discretization framework that distills a continuous foundation model into a discrete codebook, then uses it as the reconstruction target for masked image modeling (MIM). It combines gradient-free codebook learning via Growing Neural Gas, which yields an adjacency graph over the codes, with a graph-distance soft cross-entropy loss that replaces one-hot MIM targets with topologically graded supervision. With student pretraining on 383 public Duke iAMD optical coherence tomography (OCT) volumes, the pipeline distills an 86M-parameter teacher into a 22M-parameter student with at least 3.6x fewer FLOPs. The student performs close to, and in some settings above, its teacher across three classification and two segmentation benchmarks, with the magnitude and direction of the difference depending on the teacher, task, metric, and student capacity. On the external OLIVES benchmark, it surpasses its MIRAGE teacher on every metric (98.06
Fourier-domain (FD) optical coherence tomography (OCT) depends on broadband sources to maximize axial resolution and image quality. However, these lasers significantly drive device cost or may be unavailable at desired wavelength and bandwidth ranges. A potential solution lies in integrating multiple, more affordable sources with lower individual bandwidth into a single system. However, difficulties arise if the resulting spectrum exhibits discontinuities. In this letter, we present a method that can combine OCT images from a flexible number of spectra with arbitrary, possibly non-overlapping gaps using a neural network. Compared to low-resolution input images, reconstructed B-scans are super-resolved, preserve even fine and low-contrast details and edges, and exhibit strong noise reduction that increases with the band gap. The proposed method could thereby provide a major step towards high-quality OCT imaging using spectrally disjoint, low-bandwidth sources. In broadband settings, it can be directly applied as a Fourier-domain masked autoencoder for self-supervised image quality enhancement.
Background/Objectives: Diabetic macular edema is caused by disruptions in the blood-retinal barrier, resulting in leakage and inflammation, appearing in optical coherence tomography as retinal fluid, hard exudates, and hyper-reflective foci. This study aimed to assess volumetric changes in hyper-reflective foci and hard exudates in faricimab-treated diabetic macular edema in real-world practice. Methods: This retrospective longitudinal study included forty-four eyes with faricimab-treated diabetic macular edema and follow-up of ≥40 weeks. Optical coherence tomography scans were analyzed at baseline, post-loading, 6 months and 1 year. Hyper-reflective foci and hard exudates were quantified across macular subfields using a machine-learning algorithm with manual correction. Outcomes included volumetric changes, best-corrected visual acuity, and central subfield thickness. Early morphologic responders were defined post-loading by ≥20% central subfield thickness reduction and/or central subfield thickness below 280 µm. Results: At 1 year, hyper-reflective foci and hard exudate volumes significantly decreased in the central subfield (median differences: -0.025 nL, p = 0.00028; -0.119 nL, p = 0.00025) and outer ring (-0.178 nL, p = 0.00041; -0.612 nL, p = 0.014). Hard exudate volume significantly decreased in the inner ring (-0.740 nL, p < 0.0001). Best-corrected visual acuity and central subfield thickness improved significantly (-0.10 logarithm of the minimum angle of resolution (logMAR), p = 0.0018; -48 µm, p < 0.0001). Total hyper-reflective foci volumes were similar between responders and non-responders. In contrast, hard exudate burden remained numerically higher in non-responders. Conclusions: During the year following the switch to faricimab, HRF and HE burden decreased alongside visual and anatomical improvement. HRF volumes were descriptively similar between early CST responders and non-responders, whereas HE burden remained numerically higher in non-responders.
This work decomposes the wavefront at the exit pupil of an optical system into two parts: the viewpoint–invariant and the viewpoint–dependent parts. The decomposition is achieved by calculating the first and second derivatives of the wavefront and by constructing viewpoint–invariant curvature maps and viewpoint–dependent descriptors. Analysis of the former yields intrinsic properties of the wavefront, such as the radius of the base sphere, the asphericity, etc., while the latter include, among other things, an astigmatism map, the vertex and the tilt. These maps/descriptors provide intuitive information suitable for evaluation and compensation of aberrations in the optical system.
