Related Experiment Video
Updated: Nov 1, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.1K
Evaluation of Explainable Deep Learning Methods for Ophthalmic Diagnosis
Amitojdeep Singh1,2, Janarthanam Jothi Balaji3, Mohammed Abdul Rasheed1
1Theoretical and Experimental Epistemology Laboratory (TEEL), School of Optometry and Vision Science, University of Waterloo, Waterloo, ON, Canada.
Clinical Ophthalmology (Auckland, N.Z.)
|June 28, 2021
Summary
Deep Taylor, an attribution method, provided the best explanations for deep learning models in retinal OCT diagnosis, according to clinician ratings. This aids in the clinical acceptance of AI for diagnosing eye diseases.
Area of Science:
- Ophthalmology
- Artificial Intelligence
- Medical Imaging
Background:
- Deep learning models show high accuracy in medical imaging but lack explainability, hindering clinical adoption.
- Attribution methods aim to explain AI decisions, but their performance on medical images is understudied.
- This study compares attribution methods for retinal optical coherence tomography (OCT) diagnosis.
Purpose of the Study:
- To compare the clinical significance of explanations from various attribution methods for retinal OCT diagnosis.
- To identify the most effective attribution method for explaining deep learning models in this context.
Main Methods:
- Trained the Inception-v3 deep learning model to diagnose choroidal neovascularization (CNV), diabetic macular edema (DME), and drusen from retinal OCT scans.
- Evaluated explanations from 13 attribution methods using a panel of 14 clinicians.
- Collected clinician feedback on the significance and utility of the explanations.
Main Results:
- Deep Taylor, a Taylor series-based method, received the highest median clinician rating (3.85/5) for explanation quality.
- Guided backpropagation (GBP) and SHapley Additive exPlanations (SHAP) were the next highest-rated methods.
- Top methods successfully highlighted disease-specific features like fluid in CNV and edema boundaries in DME.
Conclusions:
- Deep Taylor offers the most clinically significant explanations for retinal OCT diagnosis among the methods tested.
- The optimal attribution method may vary depending on the specific medical diagnosis task.
- Clinicians showed a high degree of acceptance for AI-driven diagnostic explanations.
