Domain Adaptive Diabetic Retinopathy Grading with Model Absence and Flowing Data

TL;DR

This work introduces Online Model-gnostic Domain Adaptation (OMG-DA) for DR grading under practical clinical constraints where source data and pre-trained models are unavailable and target data arrive in a flowing, unlabeled stream. The authors propose Generative Unadversarial Examples (GUES), a data-centric approach that learns a perturbation function via a Variational Autoencoder and uses fine-grained saliency maps as pseudo-perturbation labels to generate individualized perturbations, producing generative unadversarial examples. Theoretical support is provided by two theorems linking the generative perturbation form to traditional unadversarial optimization and bounding latent inputs with saliency maps, accompanied by extensive experiments on four DR benchmarks showing consistent improvements over baselines, including when combined with online TTA methods and at small batch sizes. The method demonstrates strong interpretability, focusing perturbations on DR-related lesions, and achieves robustness under flowing-data constraints, suggesting practical utility for real-world deployment with privacy and data-flow considerations. Overall, GUES advances a data-centric, model-absent framework for domain adaptation in medical imaging with demonstrated efficacy and theoretical grounding.

Abstract

Domain shift (the difference between source and target domains) poses a significant challenge in clinical applications, e.g., Diabetic Retinopathy (DR) grading. Despite considering certain clinical requirements, like source data privacy, conventional transfer methods are predominantly model-centered and often struggle to prevent model-targeted attacks. In this paper, we address a challenging Online Model-aGnostic Domain Adaptation (OMG-DA) setting, driven by the demands of clinical environments. This setting is characterized by the absence of the model and the flow of target data. To tackle the new challenge, we propose a novel approach, Generative Unadversarial ExampleS (GUES), which enables adaptation from a data-centric perspective. Specifically, we first theoretically reformulate conventional perturbation optimization in a generative way--learning a perturbation generation function with a latent input variable. During model instantiation, we leverage a Variational AutoEncoder to express this function. The encoder with the reparameterization trick predicts the latent input, whilst the decoder is responsible for the generation. Furthermore, the saliency map is selected as pseudo-perturbation labels. Because it not only captures potential lesions but also theoretically provides an upper bound on the function input, enabling the identification of the latent variable. Extensive comparative experiments on DR benchmarks with both frozen pre-trained models and trainable models demonstrate the superiority of GUES, showing robustness even with small batch size.

Figures (12)

• Figure 1: Comparison between the OMG-DA and SFDA settings. (a) In SFDA, adaptation builds upon the accumulated data, which demands significant storage resources in the hospital. Additionally, the models' architecture and parameters are accessible, exposing them to potential attacks. (b) OMG-DA provides a practical scenario: Flowing data mimic the patients' arrival in a stream way, and the models are unseen (strictly controlled) before using them, avoiding attacks like membership inference attacks shokri2017membership.
• Figure 2: The instantiation framework of GUES in the OMG-DA setting. (a) For target input $x_t$, the VAE model generates individual perturbation $\delta_t={F}_{\Phi}\left(\frac{\partial \delta_0}{\partial x} \right)$. After that, the by-pass path incorporates $\delta_t$ and $x_t$ to create the generative unadversarial example $\hat{x}_t$. Treating $x_t$'s saliency map $g_t$ as reconstruction supervision for model training. (b) At the inference phase, the generated unadversarial example $\hat{x}_t$ is directly provided to the frozen source model or other trainable models.
• Figure 3: Explanation in choosing fine-grained saliency maps as supervision. Left: The testing fundus image selected from "Moderate DR" class in APTOS demonstrates that H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. Middle: The saliency map highlights those lesions. Right: The gradient-CAM visualization of the source model on task DDR$\to$APTOS.
• Figure 4: Comparison results with batch size varying from 2 to 64 over the 12 tasks (The details are provided in Supplementary.). Left, middle, and right report ACC, QWK, and AVG, respectively.
• Figure 5: Interpretability analysis based on a typical fundus image from "Moderate DR" class in APTOS. Here, H (hemorrhages), SE (soft exudates), and EX (hard exudates) are essential characteristics to judge the DR grade. The gradient CAM-based heatmap of five models visualizes the capture of those lesions. All models are trained on task DDR$\to$APTOS, where Oracle is trained using ground truth in APTOS.

Theorems & Definitions (2)

• Theorem 1
• Theorem 2