ADLER-MRI: Adaptive Deep Learning for Enhanced MRI Reconstruction with Noise-Resilient Models

Medical image reconstruction, particularly within the field of MRI, has witnessed a transformative advancement in recent years, primarily attributed to the integration of deep learning techniques. Nevertheless, these advanced deep reconstruction models often encounter generalization difficulties when dealing with out-of-distribution samples from unseen domains. Privacy regulations frequently restrict access to source-domain training data, complicating efforts to retrain or fine-tune models for target-domain test data—a scenario frequently encountered in clinical practice. In response to these challenges, we propose a novel method: a black-box test-time adaptation method tailored for MRI reconstruction, regardless of noise variations. The proposed method, named ADLER-MRI, is grounded in prior-informed implicit neural representation (INR) learning, which capitalizes on the power of INR to synthesize image representations across a spectrum of noise conditions. ADLER-MRI adeptly adjusts input representations by analyzing the outputs inferred from an opaque reconstruction model, at test time. Our empirical investigations, utilizing the IXI and fastMRI datasets for test-time adaptation in MRI reconstruction with variable noise conditions, demonstrate that ADLER-MRI surpasses current leading algorithms in performance.