Our lab develops advanced deep learning algorithms to analyze ocular images and extract clinically meaningful features associated with a wide range of diseases. Ocular imaging provides rich information about the structure, vasculature, and health of the eye, but many disease-related changes can be subtle and difficult to detect through visual inspection alone. By transforming complex imaging data into quantitative representations, our research aims to reveal patterns that may support earlier disease detection, more accurate diagnosis, objective risk assessment, and a better understanding of disease development and progression.
This figure illustrates an example of our computational approach. A deep learning encoder processes a color fundus photograph through multiple vision transformer blocks and converts the original image into 25 distinct feature maps. Each feature map emphasizes a different aspect of the photograph, including anatomical structures, vascular patterns, tissue appearance, color distributions, texture variations, and potential abnormalities. Some maps capture broad structural information, while others respond to fine details that may not be immediately apparent to the human eye. The algorithm combines these complementary features to create a compact and informative representation of the fundus image. We then investigate how these learned imaging features are associated with specific ocular or systemic diseases, with the ultimate goal of developing reliable and interpretable tools for clinical research and patient care.
Our lab also develops automated methods for segmenting retinal layers in optical coherence tomography (OCT) images. OCT provides high-resolution, three-dimensional views of the retina, allowing us to measure subtle structural changes that may serve as biomarkers of ocular and systemic diseases. Accurate layer segmentation enables us to quantify retinal morphology through measurements such as layer thickness, volume, and spatial variation. It also allows us to generate en-face images that capture textural and reflectance patterns within specific retinal layers. These measurements provide complementary information about retinal structure and tissue characteristics.
This figure illustrates several outputs and challenges associated with our approach. Panel (a) shows a central macular B-scan with automated segmentation of multiple retinal layer boundaries. Panel (b) demonstrates motion artifacts along the B-scan acquisition direction, which can introduce distortions in the retinal anatomy. Despite these challenges, our segmentation method remains highly accurate and robust. This performance enables reliable analysis of OCT volumes acquired under real-world conditions, including images affected by motion-related variability. From the segmented OCT volume, we can generate a thickness map of the ganglion cell–inner plexiform layer (GCIPL), as shown in panel (c), and a corresponding GCIPL en-face image, as shown in panel (d). Panels (e) and (f) present a thickness map and an en-face image, respectively, for the region extending from the photoreceptor outer segments (OS) to the retinal pigment epithelium (RPE). These quantitative and layer-specific representations can support disease detection, progression monitoring, and biomarker discovery.