Research map / Label-efficient and robust learning / Test-time adaptation

Efficient Deformable Convolutional Prompt for Continual Test-Time Adaptation in Medical Image Segmentation

AAAI 2025 · Label-efficient and robust learning › Test-time adaptation · approach: Test-time adaptation of vision-language models · 2 citations (OpenAlex)

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 EDCP, a parameter-efficient deformable convolutional visual prompt with an offset bank that skips redundant training, for continual test-time adaptation in medical image segmentation.

Problem

Continual test-time adaptation of a frozen source segmentation model to continually changing unlabeled medical domains is costly, since existing prompt methods use large or complex prompts and expensive injection strategies (uncertainty, FFT) that inflate trainable parameters and compute on high-dim

Approach

Replaces standard image-sized prompts with a deformable convolution prompt whose sampling offsets are produced by a decomposed offset branch; the decomposed second-stage offset matrices are stored in an offset bank acting as implicit domain indicators, and training is skipped for samples whose offsets are highly similar to stored ones.

Key results

• On OD/OC segmentation, EDCP reaches 69.03% average DSC versus 67.05% for VPTTA and 63.68% for Source Only.
• On BraTS 2020 cross-sequence, EDCP averages 28.76% DSC versus 15.51% for VPTTA and 24.04% for Source Only.
• Uses 186 trainable parameters with relative training time 1.5 and an average of 0.03 backward passes per sample, versus 75 parameters for VPTTA with the highest compute.

Measured effects reported in the paper

Surprising findings

• Using the deformable convolution prompt with entropy minimization alone hurts performance (61.88 DSC vs 63.68 Source Only in the ablation), so the alignment loss is needed.
• Existing CTTA baselines such as CoTTA and TENT fall below Source Only on brain tumor segmentation, indicating error accumulation under large cross-sequence shifts.

Experimental setup

Datasets: RIM-ONE-r3 REFUGE ORIGA REFUGE-Validation Drishti-GS BraTS 2020 MMWHS 2017

Benchmarks: OD/OC fundus segmentation (5 domains) BraTS 2020 brain tumor segmentation (cross-sequence) MMWHS 2017 cardiac segmentation (cross-modality)

Metrics: DSC relative training time trainable parameters GPU memory average forward/backward passes

Baselines: Source Only TENT-continual CoTTA SAR VPTTA

Builds on: VPTTA CoTTA Tent Deformable convolution Convolutional visual prompt

Closest papers by meaning

• Each Test Image Deserves A Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation (CVPR 2024 · similarity 0.91)
• Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation (AAAI 2025 · similarity 0.87)
• CAT: Coordinating Anatomical-Textual Prompts for Multi-Organ and Tumor Segmentation (NeurIPS 2024 · similarity 0.86)
• Exploring Sparse Visual Prompt for Domain Adaptive Dense Prediction (AAAI 2024 · similarity 0.85)
• Towards a Comprehensive, Efficient and Promptable Anatomic Structure Segmentation Model Using 3D Whole-Body CT Scans (AAAI 2025 · similarity 0.85)
• E2ENet: Dynamic Sparse Feature Fusion for Accurate and Efficient 3D Medical Image Segmentation (NeurIPS 2024 · similarity 0.85)