Research map / Label-efficient and robust learning / Test-time adaptation
Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GraTa, a test-time adaptation method for medical image segmentation that aligns a pseudo (consistency) gradient with an auxiliary (entropy) gradient and sets the learning rate from their cosine similarity.
Problem
Test-time adaptation for optic disc and cup segmentation under domain shift is limited by suboptimal optimization direction (pseudo gradient from self-supervised objectives) and a fixed learning rate.
Approach
Adds an auxiliary entropy gradient to the consistency-loss pseudo gradient, implicitly maximizing their inner product by evaluating the consistency loss at parameters updated by the entropy gradient, and scales the learning rate by a custom function of the cosine similarity between the two gradients.
Key results
- GraTa achieves an average DSC of 74.34 across the five domains, improving over the No Adapt baseline (68.47) by 5.87 points and over the best competing method SAR (72.67) and DeTTA (72.69).
- On target Domain D, GraTa reaches 66.84 DSC versus 54.07 for No Adapt (+12.77).
- Replacing the fixed learning rate with the cosine-similarity-based dynamic rate adds 2.81 points, and the gradient-alignment objective adds 2.14 points in ablation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GraTa instead of no adaptation | ResUNet-34 · Average across domains A–E | DSC: +5.87% | Helped |
| Add entropy-gradient alignment to the consistency objective (ablation) | GraTa with fixed β · Average across domains A–E | DSC: 69.39 -> 71.53 | Helped |
| Use dynamic learning rate η instead of fixed β (ablation) | Gradient-aligned consistency objective · Average across domains A–E | DSC: 71.53 -> 74.34 | Helped |
| Use consistency loss as pseudo objective and entropy loss as auxiliary objective (ablation) | GraTa · Average across domains A–E | DSC: 71.47 -> 74.34 | Helped |
| Replace Cus with linear activation (ablation) | GraTa learning-rate mapping · Average across domains A–E | DSC: 74.34 -> 73.71 | Hurt |
| Replace Cus with sigmoid activation (ablation) | GraTa learning-rate mapping · Average across domains A–E | DSC: 74.34 -> 73.83 | Hurt |
| Replace Cus with ReLU activation (ablation) | GraTa learning-rate mapping · Average across domains A–E | DSC: 74.34 -> 72.92 | Hurt |
| Replace Cus with Softplus activation (ablation) | GraTa learning-rate mapping · Average across domains A–E | DSC: 74.34 -> 73.32 | Hurt |
| Replace entropy auxiliary loss with reconstruction loss (ablation) | Consistency pseudo objective · Average across domains A–E | DSC: 74.34 -> 73.81 | Hurt |
| Replace entropy auxiliary loss with rotation prediction loss (ablation) | Consistency pseudo objective · Average across domains A–E | DSC: 74.34 -> 74.08 | Hurt |
| Replace entropy auxiliary loss with super-resolution loss (ablation) | Consistency pseudo objective · Average across domains A–E | DSC: 74.34 -> 74.02 | Hurt |
| Replace entropy auxiliary loss with denoising loss (ablation) | Consistency pseudo objective · Average across domains A–E | DSC: 74.34 -> 73.64 | Hurt |
Surprising findings
- Using the entropy loss as the pseudo objective with consistency as auxiliary (swapped roles) gives weaker results, and the paper reports that auxiliary losses based on reconstruction, rotation, super-resolution, or denoising help the consistency objective only as auxiliaries rather than as pseudo ob
Limitations
- Evaluated only on a single task (optic disc and cup segmentation) with five fundus datasets.
- Scaling factor beta set empirically to 0.0001 without a broader sensitivity study shown.
- Gradient alignment relies on an implicit first-order Taylor approximation that drops the remainder term.
Experimental setup
Datasets: RIM-ONE-r3 REFUGE ORIGA REFUGE-Validation/Test Drishti-GS
Benchmarks: RIM-ONE-r3 REFUGE ORIGA REFUGE-Validation/Test Drishti-GS
Metrics: Dice score (DSC)
Baselines: No Adapt DUA DIGA MedBN TENT DLTTA DomainAdaptor SAR DeTTA
Builds on: TENT PCGrad DUA SAR
Closest papers by meaning
- Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation (CVPR 2024 · similarity 0.89)
- Efficient Deformable Convolutional Prompt for Continual Test-Time Adaptation in Medical Image Segmentation (AAAI 2025 · similarity 0.87)
- Adaptive Bidirectional Displacement for Semi-Supervised Medical Image Segmentation (CVPR 2024 · similarity 0.85)
- Test-Time Adaptation via Style and Structure Guidance for Histological Image Registration (AAAI 2024 · similarity 0.84)
- Test-Time Adaptation of Vision-Language Models for Open-Vocabulary Semantic Segmentation (NeurIPS 2025 · similarity 0.84)
- Domain-Specific Block Selection and Paired-View Pseudo-Labeling for Online Test-Time Adaptation (CVPR 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.