TL;DR
This paper introduces a novel diffuser-based fundus camera that captures eye fundus images and allows for computational refocusing after image acquisition. This approach eliminates the need for mechanical focusing mechanisms, potentially leading to simpler, lower-cost, and more accessible diagnostic tools for eye diseases.
What Problem Does This Paper Address?
Comprehensive eye examinations are crucial for detecting and managing vision-threatening diseases like diabetic retinopathy and glaucoma. However, current fundus cameras, while advanced, are often expensive, complex, and require skilled operators for precise focusing. This high cost and complexity limit accessibility, especially in underserved areas. Existing portable fundus cameras, while more accessible, can still be challenging to align and focus. The paper aims to address these barriers by developing a computational imaging approach that simplifies the imaging process and enables post-capture focus adjustment, making fundus imaging more user-friendly and cost-effective.
Methods
The core of the system is a diffuser-based computational imaging technique. A holographic diffuser is placed in a modified commercial fundus camera, specifically at a plane conjugate to the eye's pupil. This diffuser encodes wavefront information from the fundus onto the image sensor. The key innovation lies in the computational reconstruction process. By measuring the system's point spread function (PSF) for various defocus errors, the captured image can be computationally deconvolved with the appropriate PSF to digitally refocus it. This allows for the reconstruction of a "focal stack" of images, from which the sharpest can be selected. The system captures three consecutive images under red, green, and blue illumination to achieve color fundus imaging. The reconstruction uses Tikhonov regularization, a computationally efficient method for solving inverse problems.
Datasets
Not discussed in the paper. The paper focuses on system development and evaluation using a USAF 1951 resolution target and in vivo human eye imaging.
Key Insights and Ideas
- Post-capture refocusing is feasible in vivo: The paper demonstrates that computational refocusing, previously shown in model eyes, can be successfully applied to living human eyes, a significant step towards practical clinical application.
- Eliminating mechanical focus simplifies design: By moving the focusing mechanism from hardware to software, the system's complexity and cost can be significantly reduced, paving the way for more accessible diagnostic tools.
- Diffuser-based imaging offers a unique advantage: The diffuser encodes wavefront information, allowing for the reconstruction of images across a range of defocus errors without physically adjusting the camera. This is particularly beneficial for patients with significant refractive errors.
- Computational imaging can integrate multiple functions: The authors suggest that this platform could potentially be extended to simultaneously perform fundus imaging and autorefraction, further streamlining eye examinations.
Key Results
- The system successfully captured and digitally refocused color in vivo human fundus images.
- The field of view (FOV) was measured to be approximately 40 degrees, comparable to standard nonmydriatic fundus cameras.
- The resolution achieved was 7.7 to 9.6 line pairs per millimeter (lp/mm) across a range of tested defocus errors (-3D to +12D). While lower than conventional cameras (which can achieve >50 lp/mm), this resolution remained consistent without mechanical adjustments.
- The computational refocusing process was demonstrated to improve the visibility and contrast of blood vessels in the fundus images.
- The device was evaluated for light safety and found to be non-hazardous for human use according to ANSI Z80.36-2021 standards.
Limitations and Future Work
- Resolution: The current resolution (7.7-9.6 lp/mm) is lower than conventional fundus cameras. Future work could involve using encoding elements with sparser PSFs, such as random microlens arrays, or exploring advanced deep learning-based deconvolution methods to improve resolution.
- Alignment: The lack of a fixation target made alignment challenging, and future iterations could incorporate such features.
- Illumination Uniformity: The modified illumination system resulted in some unevenness and reduced contrast in conventional images compared to clinical standards.
- PSF Optimization: Further optimization of the PSF could improve the resolution of computational images.
- Integration of Aberrometry: The authors suggest future work could synergistically estimate aberrometry information (for eyeglass prescriptions) and reconstruct a fundus image from a single measurement, creating a combined diagnostic tool.
Summary updated