• AI, Transfer Learning, and Retinal Disease

    Written By: Jean Shaw
    Selected By: Deepak P. Edward, MD

    Journal Highlights


    Download PDF

    Artificial intelligence (AI) systems typically employ a highly specialized deep learning machine and a dataset of millions of images. Kermany et al. evaluated a new deep learning frame­work that uses transfer learning, thus allowing these systems to use a smaller dataset of images. They found that their system effectively classified spectral-domain optical coherence tomography (SD-OCT) images of age-related mac­ular degeneration and diabetic macular edema (DME), matching the proficien­cy of human experts.

    For the study, a dataset of 108,312 SD-OCT images from 4,686 patients was used to train the deep learning framework. The model was then tested with a validation dataset of 1,000 im­ages from 633 patients, with the images evenly drawn from image subsets of choroidal neovascularization (CNV), DME, drusen, and no disease.

    The AI system categorized the OCT images as “urgent referrals” (those with CNV or DME); “routine referrals” (those with drusen); and “observation” (those with no disease), achieving an accuracy rate of 96.6%, with a sensitiv­ity of 97.8% and specificity of 97.4%. An independent test set of images was used to compare the network’s referral decisions with those made by 6 experi­enced ophthalmologists; the network’s performance was comparable to that of the human experts.

    The researchers also performed occlusion testing to identify the areas of greatest importance used by their AI system in assigning a diagnosis. They noted that the greatest benefit of occlu­sion testing is that it sheds light on how neural networks “think,” thus making the process more transparent and bol­stering confidence in the results. In this study, the occlusion tests confirmed that the AI system made its decisions using accurate distinguishing features.

    In a novel twist, the researchers also used their system to evaluate chest x-ray images for the purposes of diagnosing pediatric pneumonia. They found that the system successfully differentiated between viral and bacterial pneumo­nia, with an accuracy of 92.8%. This demonstrates that the system can be applied to a wide range of medical imaging techniques across multiple medical specialties, they said.

    The original article can be found here.