Many of today’s most advanced AI models require enormous amounts of expertly annotated images to learn how to identify disease. That challenge becomes even greater when working with three-dimensional scans, where each image can contain dozens or even hundreds of individual slices that must be reviewed by highly trained clinicians.

A study co-authored by Houston Methodist retina specialist Dr. Charles Wykoff offers a potential solution.

Published in Nature Biomedical Engineering, the research describes a deep learning framework called SLIViT (Slice Integration by Vision Transformer) that can accurately analyze volumetric medical imaging data while requiring far fewer annotated training examples than traditional approaches. The model was evaluated across multiple imaging modalities, including optical coherence tomography (OCT), ultrasound, MRI and CT scans.

What makes the approach notable is its ability to leverage knowledge from large collections of two-dimensional medical images and apply it to more complex three-dimensional scans.

“Obtaining expert annotations for volumetric imaging studies is often labor intensive and expensive. By building on existing 2D image datasets, the model helps overcome one of the biggest bottlenecks facing AI development in medical imaging.”

---

Dr. Charles Wykoff, retina specialist

For retina specialists, the findings are particularly relevant. OCT scans have become foundational tools for diagnosing and monitoring retinal diseases such as age-related macular degeneration (AMD). Yet interpreting these scans remains a time-intensive process that requires careful review by experienced clinicians.

In testing, SLIViT demonstrated strong performance in identifying several high-risk AMD biomarkers from 3D OCT scans, outperforming existing domain-specific AI models. The system was also tested against human graders and achieved performance comparable to clinical specialists while completing analyses dramatically faster.

The implications extend well beyond ophthalmology.

The investigators successfully applied the same framework to cardiac ultrasound videos, liver MRI scans and chest CT scans, suggesting the approach may provide a versatile foundation for a wide range of imaging applications. Rather than building a new AI architecture for every specialty and imaging modality, researchers may be able to adapt a common framework across multiple clinical domains.

According to Dr. Wykoff, this could help accelerate research involving newly discovered disease biomarkers. Once experts annotate a relatively small training set, AI systems may be able to scale that knowledge across thousands of additional scans, reducing both time and cost while preserving specialist expertise.

The technology is not intended to replace clinicians. Instead, researchers envision AI serving as a force multiplier — helping physicians process increasingly complex imaging datasets, identify important findings more efficiently and focus their attention where it matters most.