Adversarial Multi-task Learning for Joint Detection of Viral Retinitis using Common Blood Test and Serology Test Data
Purpose
Multi-task learning (MTL) models were developed with adversarial training to accurately detect ARN, CMV retinitis, and non-infectious uveitis of the posterior segment (NIU-PS) with hematological data.
Methods
This cross-sectional study included 3080 eyes. Single-task learning models were developed to detect each disease respectively, followed by diverse MTL models for the joint detection of 3 diseases. Model performance was evaluated in terms of area under the ROC curve and area under the PR curve. Task-specific important features and model trustworthiness were investigated via SHAP and the counterfactual inference test.
Results
Among the 4 deep learning models, adversarial MTL yielded the highest classification performance and average precision for ARN detection (AUROC=0.932) and CMV retinitis detection (AUROC=0.982), respectively. The adversarial MTL model identified the 6 most important features for ARN and CMV retinitis detection. Adversarial MTL had much fewer instances of minimal prediction variation when significant input features were altered compared to fully-shared MTL, manifesting that adversarial MTL better reflects the changes in inputs.
Conclusion
MTL can be an adjuvant method to differentiate necrotizing viral retinitis from NIU-PS, allowing early antiviral therapy and preservation of functional vision. Compartmentalizing shared-private feature spaces and introducing adversarial training, adversarial MTL achieved superior performance, expanding our knowledge of the diagnostic value of common blood and viral serology tests. The counterfactual inference tests demonstrated that the classification of adversarial MTL is more reliable than fully-shared MTL.
Conflict of interest
No
Authors 1
Last name
CHOI
Initials of first name(s)
EY
Department
Department of Ophthalmology, Institute of Vision Research, Gangnam Severance Hospital, Yonsei Univer
City
Seoul
Country
Korea, Republic of
Authors 2
Last name
Kai Tzu-iunn
Initials of first name(s)
Ong
Department
Department of Artificial Intelligence, Yonsei University College of Computing
City
Seoul
Country
Korea, Republic of
Authors 3
Last name
Taeyoon
Initials of first name(s)
Kwon
Department
Department of Artificial Intelligence, Yonsei University College of Computing
City
Seoul
Country
Korea, Republic of
Authors 4
Last name
Jinyoung
Initials of first name(s)
Yeo
Department
Department of Artificial Intelligence, Yonsei University College of Computing
City
Seoul
Country
Korea, Republic of
This website uses cookies to ensure you get the best experience on our website.
Learn more