Prashant Tailor, MD
Ophthalmology Resident
Mayo Clinic
Dr. SriniVas R. Sadda, MD presented on Deep Learning for Biomarkers in Non-neovascular AMD. Dr. Sadda first discussed how OCT has largely replaced fundus photography in terms of evaluation of AMD with OCT biomarkers having been shown as validated predictors in AMD progression such as hyperreflective foci. Next Dr. Sadda postulated the question: Can we utilize deep learning to automatically identify biomarkers for AMD progression?
He highlighted a key issue with deep learning in Ophthalmology is the lack of high quality training data as it is extremely resource intensive to collect/curate large cohorts of training data. Dr. Sadda discussed how transfer learning is essential to address the issue as it gives the model a running start. Transfer learning works by borrowing information about the structure and parameters of the network from publicly available large datasets. However, he noted when the data is 3-D volumes, transfer learning cannot be directly applied unless other 3-D volumes are available (which currently there is not). To both leverage external datasets of 2-D images and counter the loss of information from converting 3-D to 2-D, he and his group developed SILVER-net (Slice Integration of Volmetric features Extracted by pre trained Residual neural networks) with UCLA computation medicine. SILVER-net takes the 3D OCT volumes and converts them to 2D “tilting” of slices allowing use of transfer learning with currently available 2-D data. SILVER-net utilizes additional layers of the deep neural network to preserve the 3-D spatial structures lost by tilting.
Using this approach, they were able to train their model despite a limited number of manually annotated examples. SILVER-net was compared to current state-of-the-art deep learning methods (3-D CNN and 2-D CNN) and they found SILVER-net to outperform previous approaches and even annotators. Utilizing attention maps, they noted the model failed when things were close like when intraretinal hyper reflective foci were close to the RPE. Overall, Dr. Sadda noted deep learning can automatically identify biomarkers for AMD progression and transfer learning is critical given the difficult to find high-quality annotated data.
