DP-Morph: Improving the Privacy-Utility-Performance Trade-off for Differentially Private OCT Segmentation

Optical Coherence Tomography (OCT) images show a cross-section of the retina and are used for early detection of retinal diseases and glaucoma through analysis of the retinal layers. Advances in deep learning have enabled state-of-the-art OCT segmentation models to support this analysis. However, using medical data to train these models raises concerns about patient privacy. For example, membership inference attacks allow an adversary to determine whether a particular data point was included in the training data. Differentially Private Stochastic Gradient Descent (DPSGD) improves the privacy of deep learning models by ensuring that these models do not disclose sensitive information about individual data points. However, implementing DPSGD may cause decreased model accuracy and/or increased computational demands. In this paper, we evaluate the privacy, utility, and computational performance of five OCT segmentation models trained using DPSGD on graphics processing units (GPUs). To improve utility, we then propose DP-Morph, a novel privacy-preserving modification of DPSGD based on morphology. We show that DP-Morph improves segmentation performance, for example, increasing the Dice coefficient of LFUNet from 0.50 to 0.70 for a privacy budget of 200.