Investigative Ophthalmology & Visual Science Cover Image for Volume 65, Issue 7
June 2024
Volume 65, Issue 7
Open Access
ARVO Annual Meeting Abstract  |   June 2024
Attention-Based Biomarker Activation Map: an Interpretability Method Specifically Designed for Deep-Learning-Aided Eye Disease Diagnosis
Author Affiliations & Notes
  • Pengxiao Zang
    Oregon Health & Science University Casey Eye Institute, Portland, Oregon, United States
  • Tristan T Hormel
    Oregon Health & Science University Casey Eye Institute, Portland, Oregon, United States
  • Thomas S Hwang
    Oregon Health & Science University Casey Eye Institute, Portland, Oregon, United States
  • Yali Jia
    Oregon Health & Science University Casey Eye Institute, Portland, Oregon, United States
  • Footnotes
    Commercial Relationships   Pengxiao Zang None; Tristan Hormel None; Thomas Hwang None; Yali Jia Genentech, Inc., Code F (Financial Support), Genentech, Inc., Code P (Patent), Optovue/Visionix, Inc., Code P (Patent), Optos, Code P (Patent), Optovue/Visionix, Inc., Code R (Recipient)
  • Footnotes
    Support  National Institute of Health (R01 EY027833, R01 EY035410, R01 EY024544, R01 EY031394, T32 EY023211, UL1TR002369, P30 EY010572); the Malcolm M. Marquis, MD Endowed Fund for Innovation; an Unrestricted Departmental Funding Grant and Dr. H. James and Carole Free Catalyst Award from Research to Prevent Blindness (New York, NY), Edward N. & Della L. Thome Memorial Foundation Award, and the Bright Focus Foundation (G2020168, M20230081).
Investigative Ophthalmology & Visual Science June 2024, Vol.65, 1589. doi:
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Pengxiao Zang, Tristan T Hormel, Thomas S Hwang, Yali Jia; Attention-Based Biomarker Activation Map: an Interpretability Method Specifically Designed for Deep-Learning-Aided Eye Disease Diagnosis. Invest. Ophthalmol. Vis. Sci. 2024;65(7):1589.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose : Deep-learning-aided classifiers can achieve specialist-level performances in several eye disease diagnostic tasks. However, the output must be explainable for integration into clinical workflow. We propose a novel attention-based biomarker activation map (AttnBAM) generation framework that highlights clinically important biomarkers that make the classifier output intuitively interpretable for the clinician.

Methods : We trained an age-related macular degeneration (AMD) classifier for color fundus photography (CFP) images and a referable diabetic retinopathy (DR) classifier for optical coherence tomography angiography (OCTA). The AMD CFP data set was a publicly available data set comprising 500 healthy and 326 clinically evident AMD eyes. A CFP image of the macular region was the input for the AMD classifier. The OCTA data set comprised 257 non-referable DR and 199 referable DR eyes graded based corresponding color photographs from a longitudinal study of DR. A 3×3-mm macular superficial vascular complex en face image was the input for the DR classifier. We trained an AttnBAM generation framework for each classifier. The generator in the framework was designed by combining a transformer-based encoder and a convolution-based decoder (Fig. 1). The main generator was mainly trained to produce an output image that the classifier would classify as negative from a positive input, while the assistant generator was mainly trained to do the opposite. The AttnBAM was then constructed as the absolute difference image between the positive input and negative output of the main generator. Data was divided into training (60%), validation (20%), and testing (20%) sets in the development of both classifiers and generation frameworks.

Results : The AMD and referable DR diagnostic classifiers achieved area under the receiver operating characteristic curve (AUC) of 0.85 and 0.97, respectively. The generated AttnBAMs accurately and sharply highlighted biomarkers used by the AMD and DR classifiers, respectively (Fig. 2), highlighting choroidal neovascularization on the AMD CFP and non-perfusion area on the referable DR OCTA.

Conclusions : The generated AttnBAMs highlighted clinically relevant biomarkers in the two modalities sufficient to aid clinicians utilize deep learning classifiers in AMD and DR. This innovation may also facilitate new biomarker discovery with CFP and OCTA.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

 

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×