Quantitative and Qualitative Evaluation of Explainable Deep Learning\n Methods for Ophthalmic Diagnosis

Abstract

Background: The lack of explanations for the decisions made by algorithms\nsuch as deep learning has hampered their acceptance by the clinical community\ndespite highly accurate results on multiple problems. Recently, attribution\nmethods have emerged for explaining deep learning models, and they have been\ntested on medical imaging problems. The performance of attribution methods is\ncompared on standard machine learning datasets and not on medical images. In\nthis study, we perform a comparative analysis to determine the most suitable\nexplainability method for retinal OCT diagnosis.\n Methods: A commonly used deep learning model known as Inception v3 was\ntrained to diagnose 3 retinal diseases - choroidal neovascularization (CNV),\ndiabetic macular edema (DME), and drusen. The explanations from 13 different\nattribution methods were rated by a panel of 14 clinicians for clinical\nsignificance. Feedback was obtained from the clinicians regarding the current\nand future scope of such methods.\n Results: An attribution method based on a Taylor series expansion, called\nDeep Taylor was rated the highest by clinicians with a median rating of 3.85/5.\nIt was followed by two other attribution methods, Guided backpropagation and\nSHAP (SHapley Additive exPlanations).\n Conclusion: Explanations of deep learning models can make them more\ntransparent for clinical diagnosis. This study compared different explanations\nmethods in the context of retinal OCT diagnosis and found that the best\nperforming method may not be the one considered best for other deep learning\ntasks. Overall, there was a high degree of acceptance from the clinicians\nsurveyed in the study.\n Keywords: explainable AI, deep learning, machine learning, image processing,\nOptical coherence tomography, retina, Diabetic macular edema, Choroidal\nNeovascularization, Drusen\n

Ähnliche Arbeiten