Susceptibility to misdiagnosis of adversarial images by deep learning based retinal image analysis algorithms

Deep learning algorithms, typically implemented as Convolutional Neural Networks (CNNs), in recent years have gained traction in medical image analysis. The majority of CNNs employed in retinal image diagnosis applications are image-based; wherein input is the retinal image and output is the classification/diagnosis, resulting in a blackbox like algorithm. In contrast, hybrid lesion-based algorithms employ multiple CNN-based detectors to categorically detect various lesions in the image, and final diagnosis is computed from combination of detector outputs. Such algorithms are more physiologically plausible and provides explainability of the final prediction through intermediate detector outputs. Both classes of algorithms have reported equal diagnostic performance and outperform clinical experts in detecting referable diabetic retinopathy (rDR). However, CNNs are sensitive to adversarial images where a limited number of pixels are modified by a fraction of intensity while preserving global image context; leading to CNN misclassification. We compared diagnostic accuracy of the two classes of algorithms on adversarial images generated from rDR retinal images and results show that image-based CNNs are significantly more susceptible to adversarial images than hybrid lesion-based algorithms.