TLDR: This research evaluates deep learning models, particularly conditional diffusion models (cDDPMs), for improving the quality of vitreous Optical Coherence Tomography (OCT) images and reducing acquisition time. The study found that cDDPMs generate clinically meaningful high-quality images, achieving a nearly fourfold speedup compared to traditional methods, and preserving anatomical details well, as assessed by expert ophthalmologists in Visual Turing Tests. The findings emphasize the importance of clinical evaluation over purely quantitative metrics for medical imaging AI.
Optical Coherence Tomography (OCT) is a vital non-invasive imaging technique in ophthalmology, offering detailed views of the eye’s internal structures. However, capturing high-quality images of the vitreous body, the transparent gel-like substance filling the eye, presents significant challenges. Its transparency makes it difficult to visualize, and common issues like speckle (a granular pattern) and patient motion during acquisition can obscure fine anatomical details, reducing diagnostic clarity.
Traditional methods to enhance OCT images, such as signal averaging, involve taking multiple scans and averaging them. While effective, this dramatically increases acquisition time, burdening patients and reducing clinical efficiency. Other techniques like filtering and Bayesian methods often blur crucial details or require extensive tuning. Deep learning models, including Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), have shown promise but can suffer from unstable training, introduce artifacts, or cause blurriness.
This research explores the potential of deep generative models, particularly diffusion models, to overcome these limitations and enhance vitreous OCT image quality while reducing acquisition time. The study compared Conditional Denoising Diffusion Probabilistic Models (cDDPMs) and Brownian Bridge Diffusion Models (BBDMs) with established deep learning architectures like U-Net, Pix2Pix, and Vector-Quantised Generative Adversarial Network (VQ-GAN).
The goal was to generate high-quality spectral-domain (SD) vitreous OCT images from lower-quality inputs (ART10 images), aiming for a quality level similar to “pseudoART100” images, which are typically obtained by averaging ten ART10 images. A key innovation was the development of a weighted image averaging method to create the ground truth pseudoART100 images, effectively handling motion artifacts that appear as black strips in the input scans.
Model performance was rigorously assessed using both quantitative image quality metrics and, crucially, qualitative evaluations through Visual Turing Tests conducted by expert ophthalmologists. Quantitative metrics included Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS).
The results revealed interesting discrepancies between quantitative scores and clinical judgment. U-Net achieved the highest PSNR and SSIM, indicating good pixel-wise similarity. However, for perceptual quality (LPIPS), Pix2Pix and cDDPM performed better. In the first Visual Turing Test, where ophthalmologists ranked generated images, cDDPM was rated highest among the deep learning models, closely followed by U-Net. Clinicians, on average, still preferred the “true” signal-averaged high-quality images, but cDDPM and U-Net were not significantly different from this gold standard in clinical judgment.
A second, more detailed Visual Turing Test focused on the best-performing model, cDDPM. It achieved a 32.9% “fool rate,” meaning ophthalmologists mistakenly identified the generated image as the real high-quality image about a third of the time. More importantly, cDDPM demonstrated an 85.7% overall anatomical preservation, with specific vitreous structures like the posterior vitreous membrane showing 84.3% preservation. Retinal layers were perfectly preserved at 100%.
One of the most significant findings is the translational relevance: cDDPMs show strong potential for clinical integration by substantially reducing image acquisition time. The model could generate a high-quality OCT image, comparable to an ART100, in approximately 2 minutes and 36 seconds (including ART10 acquisition time), a nearly fourfold speedup compared to the 10 minutes typically required for an original ART100 scan. Furthermore, the cDDPM performed comparably well when conditioned on even lower-quality ART1 images, suggesting potential for even greater time savings in the future.
The study highlights a critical point: relying solely on quantitative metrics can be misleading. The clinical evaluation by ophthalmologists proved essential in identifying models that preserve anatomical details and avoid introducing unrealistic artifacts, which is paramount for clinical safety and reliability. Models like Pix2Pix and VQ-GAN, despite sometimes achieving good LPIPS scores, were rated poorly by clinicians due to artifacts and lack of detail.
Also Read:
- Advancing Medical Image Analysis with Adaptable Foundation Models
- Unlocking Reliable AI in Medical Imaging with Limited Data: A New Theoretical Framework
While promising, the study acknowledges limitations, including a relatively small dataset of healthy subjects, which may affect generalizability to pathological cases. Future work will focus on larger, more diverse datasets, exploring pathological conditions, and developing new quantitative metrics that better align with clinical relevance. The dataset and code for this research will be made publicly available, fostering further advancements in this field. For more details, you can access the full research paper here.
