/ /

  • linkedin
  • Increase Font
  • Sharebar

    Here come the machines: Progress or replacing MDs?

    Medicine entering new dawn with artificial intelligence that may change physician dynamic


    Machine learning in ophthalmology

    As their paper in JAMA, titled “Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs,” a Google team used an AI algorithm to interpret and grade fundus photographs with various stages of diabetic retinopathy as accurately as a cohort of ophthalmologists.

    The algorithm diagnosis was compared to the majority decision of at least seven board certified ophthalmologists grading over 11,000 color fundus photos. The algorithm attained sensitivity of 97.5% and 96.1% with specificity 93.4% and 93.9% in two image sets. Using an 8% prevalence of referable diabetic retinopathy, these results yield a negative predictive value of 99.6% to 99.8%.


    Advanced network

    This Google deep-learning algorithm is an advanced artificial neural network, composed of many simple, highly interconnected processors. The nodes or processors within the system make simple calculations that are weighted and added together to produce the final output. 

    The Google system was trained using about 120,000 color fundus photos diagnosed by ophthalmologists. In the training phase, the system made a diagnostic “guess” on each image.
    It then compared its answer to the ophthalmologists’ labeled answer and adjusts the algorithm, learning how to compute the lowest possible diagnostic error. It does this again and again, hundreds of thousands of times.

    After the training was completed, the algorithm was validated in the study. About 11,000 never-before-seen images (out of sample) were shown to the algorithm, with the results compared to board certified ophthalmologists, which yielded impressive results.

    While Google’s algorithm may not be the first to have success interpreting diabetic retinopathy images, it is the most extensive and thorough, considering the sheer number of images and the fact that every image was reliably pre-labeled by ophthalmologists.


    Screening machines

    As history has shown, an initial response to machine-learning advances in medicine can be one of concern that they will replace physicians.

    It is quite true that significant disruptive technology can cause reorganizations in the workplace and physicians are not immune. As physicians, we should look critically into the future to how these assistive technologies will affect us and make adjustments accordingly.

    Consider this dire prediction in diagnostic radiology: “They should stop training radiologists now,” said Gaeoffrey Hinton, an AI computer scientist, University of Toronto. Fortunately, this is an extreme statement and does not directly apply to ophthalmology because of the procedural nature of our specialty.

    At this point, it is important to understand that these revolutionary technologies were developed to aid clinicians and are neither intended to nor will they replace physicians.
    Given that diabetes is one of the fastest growing and leading causes of blindness worldwide, the investigators at Google identified this significant unmet need that physicians and the current health-care system are not fulfilling.

    A widespread deployment of a deep learning-based, diabetic retinopathy-screening program would lower the barriers of access to areas where an eye care provider may not be present. It would determine which patients have pathology, referring on a grander scale, and allow physicians to see more patients with pathology and less healthy patients.

    In the long run, this would provide earlier detection of referable diabetic eye disease and decrease overall healthcare costs.

    The introduction of assistive, or “smart,” screening programs are likely to increase, not decrease, the volume of diabetic eye referrals-patients with real disease that need an ophthalmologist’s expertise-as a result of capturing a greater portion of the afflicted patient population with diabetic retinopathy. Clearly, they will increase the pathologic workload sent to ophthalmologists for treatment and management, thus increasing efficiency of time and of the health-care system.

    Machines better?

    Peter A. Karth, MD
    Peter A. Karth, MD, is clinical instructor at Stanford University Department of Ophthalmology, Stanford, California, USA. He may ...

    New Call-to-action

    1 Comment

    You must be signed in to leave a comment. Registering is fast and free!

    All comments must follow the ModernMedicine Network community rules and terms of use, and will be moderated. ModernMedicine reserves the right to use the comments we receive, in whole or in part,in any medium. See also the Terms of Use, Privacy Policy and Community FAQ.

    • UBM User
      In general I agree with Dr. Karth and as a consequence, the digitalization of retinal images that will allow physicians to manage individual risk prediction by monitoring digitized results of image interpretation that will enable the monitoring in individuals of digitized variable changes chronologically of which Dr. Eric Tool has enumerated. Also the immediate delivery of results of imaged screening will significantly improve follow-up outcomes as it has been demonstrated that delayed communication of the results to the patient results in 80% failure to follow-up. Certainly imaging provides much improved recognition of lesions than provider examination and should be done without pupil dilation as this severely reduces "compliance" with recommended screening. However retinal images vary tremendously in the information about the disease process that is revealed, currently based on very old lesion knowledge, and machine learning, whatever the methodology is based on the imaged aspects. Non-myd cameras that provide color fundus photographs are poor and the images presented are not correlated with vision or vision prognosis and themselves are too expensive. SLO's and OCT's, while able to image through undiluted pupils, are far too expensive and provide (thus far) image analysis outcomes that also are similarly poor. What are needed are newly derived outcomes that examine the neuronopathy of the diabetes process and interpretation of these enhanced images by experts that can be digitized and crowd sourced to provide the screening that is needed.


    View Results