AADG: Automatic Augmentation for Domain Generalization on Retinal Image Segmentation

Abstract

Convolutional neural networks have been widely applied to medical image segmentation and have achieved considerable performance. However, the performance may be significantly affected by the domain gap between training data (source domain) and testing data (target domain). To address this issue, we propose a data manipulation based domain generalization method, called Automated Augmentation for Domain Generalization (AADG). Our AADG framework can effectively sample data augmentation policies that generate novel domains and diversify the training set from an appropriate search space. Specifically, we introduce a novel proxy task maximizing the diversity among multiple augmented novel domains as measured by the Sinkhorn distance in a unit sphere space, making automated augmentation tractable. Adversarial training and deep reinforcement learning are employed to efficiently search the objectives. Quantitative and qualitative experiments on 11 publicly-accessible fundus image datasets (four for retinal vessel segmentation, four for optic disc and cup (OD/OC) segmentation and three for retinal lesion segmentation) are comprehensively performed. Two OCTA datasets for retinal vasculature segmentation are further involved to validate cross-modality generalization. Our proposed AADG exhibits state-of-theart generalization performance and outperforms existing approaches by considerable margins on retinal vessel, OD/OC and lesion segmentation tasks. The learned policies are empirically validated to be model-agnostic and can transfer well to other models. The source code is available at https://github.com/CRazorback/AADG.

Failure to Achieve Domain Invariance With Domain Generalization Algorithms: An Analysis in Medical Imaging

“…Due to the well-understood differences between domains in medical imaging, data augmentation remains a strong contender [5]. Most methods either utilize a large number of transformations [29] or introduce targeted augmentations that mimic domain differences, with handcrafted or adversarial augmentation [9]. Other methods have been introduced that rely on regularizing deep learning models, either by regularizing the feature space itself (e.g., linear dependency regularization [30] ) or by regularizing the model's parameters, done through explicit means [4] or meta-learning [31].…”

“…While there are many methods for domain generalization, most operate by attempting to address the first point of failure of generalization: Using domain specific representations [4], [5]. This can be addressed explicitly, with something as simple as data augmentation [6] that mimics the distributional shift between domains, or it can be learnt explicitly with adversarial learning [7]- [9], which directly optimizes a neural network to remove all domain identifiable information from the feature representations. In a similar vein, there are methods that use disentangled representations [10], [11] (like those used for style transfer networks) that separate domain variant and domain invariant representations, which are often learnt adversarially as well.…”

Toward hippocampal volume measures on ultra-high field magnetic resonance imaging: a comprehensive comparison study between deep learning and conventional approaches

“…The results evidently showed the robustness of DL with respect to ultra-high field MRIs across scanning acquisition protocols. The performance of deep learning models is largely influenced by the training datasets (Zhang et al, 2020;Guan et al, 2021;Lyu et al, 2022). In order to achieve optimal performance, it is crucial for testing images to follow a similar distribution to that of the training images.…”

An In-Depth Statistical Review of Retinal Image Processing Models from a Clinical Perspective

“…The work in [6] addressed the domain gap between training and testing data by introducing a novel approach for domain generalization in medical picture segmentation. In order to align retinal Optical Coherence Tomography (OCT) B-scans and account for stochastic shifts and orientation changes in retinal layers, the researchers suggested a preprocessing step in [7].…”