Single Image Test-Time Adaptation via Multi-View Co-Training

2025-06-30

AI Summary

Test-time adaptation (TTA) for medical image segmentation typically requires substantial target-domain data, hindering per-patient real-time inference in clinical settings. Method: We propose the first TTA method tailored to single 3D medical images, leveraging an uncertainty-guided self-training framework with patch-wise multi-view collaboration. It enforces both feature-level and prediction-level consistency constraints to fully exploit 3D volumetric structural information, requiring no target-domain annotations and enabling plug-and-play integration into mainstream frameworks (e.g., nnUNet). Contribution/Results: Evaluated on three public breast MRI datasets, our method achieves an average Dice coefficient improvement of 3.75% over current state-of-the-art methods, approaching the performance upper bound of full supervision. This advancement significantly enhances the clinical feasibility of TTA for 3D medical imaging by enabling efficient, annotation-free, patient-specific adaptation.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Uncertainty Representations

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: User modeling for targeted and personalized online advertising

Abstract

Test-time adaptation enables a trained model to adjust to a new domain during inference, making it particularly valuable in clinical settings where such on-the-fly adaptation is required. However, existing techniques depend on large target domain datasets, which are often impractical and unavailable in medical scenarios that demand per-patient, real-time inference. Moreover, current methods commonly focus on two-dimensional images, failing to leverage the volumetric richness of medical imaging data. Bridging this gap, we propose a Patch-Based Multi-View Co-Training method for Single Image Test-Time adaptation. Our method enforces feature and prediction consistency through uncertainty-guided self-training, enabling effective volumetric segmentation in the target domain with only a single test-time image. Validated on three publicly available breast magnetic resonance imaging datasets for tumor segmentation, our method achieves performance close to the upper bound supervised benchmark while also outperforming all existing state-of-the-art methods, on average by a Dice Similarity Coefficient of 3.75%. We publicly share our accessible codebase, readily integrable with the popular nnUNet framework, at https://github.com/smriti-joshi/muvi.git.

Problem

• Adapting trained models to new domains with single test images
• Leveraging volumetric data for medical image segmentation
• Achieving real-time inference in clinical settings without large datasets

Innovation

• Patch-Based Multi-View Co-Training for adaptation
• Uncertainty-guided self-training for consistency
• Single-image volumetric segmentation in target domain