Physics-Driven Autoregressive State Space Models for Medical Image Reconstruction

Abstract

Medical image reconstruction from undersampled acquisitions is an ill-posed inverse problem requiring accurate recovery of anatomical structures from incomplete measurements. Physics-driven (PD) network models have gained prominence for this task by integrating data-consistency mechanisms with learned priors, enabling improved performance over purely data-driven approaches. However, reconstruction quality still hinges on the network's ability to disentangle artifacts from true anatomical signals-both of which exhibit complex, multi-scale contextual structure. Convolutional neural networks (CNNs) capture local correlations but often struggle with non-local dependencies. While transformers aim to alleviate this limitation, practical implementations involve design compromises to reduce computational cost by balancing local and non-local sensitivity, occasionally resulting in performance comparable to CNNs. To address these challenges, we propose MambaRoll, a novel physics-driven autoregressive state space model (SSM) for high-fidelity and efficient image reconstruction. MambaRoll employs an unrolled architecture where each cascade autoregressively predicts finer-scale feature maps conditioned on coarser-scale representations, enabling consistent multi-scale context propagation. Each stage is built on a hierarchy of scale-specific PD-SSM modules that capture spatial dependencies while enforcing data consistency through residual correction. To further improve scale-aware learning, we introduce a Deep Multi-Scale Decoding (DMSD) loss, which provides supervision at intermediate spatial scales in alignment with the autoregressive design. Demonstrations on accelerated MRI and sparse-view CT reconstructions show that MambaRoll provides improved performance relative to state-of-the-art CNN-, transformer-, and SSM-based methods.

Generative AI for Synthetic Medical Image Generation in Oncology: Addressing Data Scarcity in AI-Driven Cancer Diagnosis
Ohmini Krishnamurthy Rajendran
International Journal of Drug Delivery Technology5 Jun 2026Peer-reviewedLicence not statedFull text20 citedEpidemiology & public healthOpenAlex
AI-driven quality enhancement of medical datasets: a closed-loop framework of synthesis, annotation, quality control, and integration
Tao Tian, Siyuan Zhang, Xuchen Yang +1
Frontiers in medicine19 Oct 2026Peer-reviewedCC BYFull textGeneral medicineEurope PMC
Precision Cervical Reconstruction: A State-Dependent Operative Framework for Cervical Insufficiency
Yong-Jin Park, Moon-Il Park
Diagnostics (Basel, Switzerland)1 Oct 2026Peer-reviewedCC BYFull textGeneral medicineEurope PMC
The Current State of Anterior Cruciate Ligament Reconstruction: Graft Selection, Lateral Augmentation, Rehabilitation, and Return-to-Play
Jeffrey George, Andrew J. Moyal, Isha K. Shah +4
Current reviews in musculoskeletal medicine14 Oct 2026Peer-reviewedCC BYFull textSurgery & anaesthesiaEurope PMC
Fornix Reconstruction in Cicatricial Ocular Surface Disease: A Narrative Review of Graft Materials, Surgical Techniques, and Medical Management.
Muhammad Awan, Cuneyt Ozkardes, Zhenyang Zhao +1
Cureus20 Sep 2026Peer-reviewedCC BYFull textMedical imaging & radiologyEurope PMC

Cite

BibTeX RIS

BibTeX

@article{kabas2026physicsdriven,
  title = {{Physics-Driven Autoregressive State Space Models for Medical Image Reconstruction}},
  author = {Bilal Kabaş and Fuat Arslan and Valiyeh Ansarian Nezhad and Kader K. Oguz and Şaban Öztürk and Emine Ülkü Sarıtaş and Tolga Çukur},
  journal = {IEEE Transactions on Medical Imaging},
  year = {2026},
  month = jul,
  doi = {10.1109/tmi.2026.3716153},
  url = {https://doi.org/10.1109/tmi.2026.3716153},
  note = {CC BY},
}

RIS

TY  - JOUR
TI  - Physics-Driven Autoregressive State Space Models for Medical Image Reconstruction
AU  - Bilal Kabaş
AU  - Fuat Arslan
AU  - Valiyeh Ansarian Nezhad
AU  - Kader K. Oguz
AU  - Şaban Öztürk
AU  - Emine Ülkü Sarıtaş
AU  - Tolga Çukur
PY  - 2026
DA  - 2026/07/23
JO  - IEEE Transactions on Medical Imaging
DO  - 10.1109/tmi.2026.3716153
UR  - https://doi.org/10.1109/tmi.2026.3716153
AB  - Medical image reconstruction from undersampled acquisitions is an ill-posed inverse problem requiring accurate recovery of anatomical structures from incomplete measurements. Physics-driven (PD) network models have gained prominence for this task by integrating data-consistency mechanisms with learned priors, enabling improved performance over purely data-driven approaches. However, reconstruction quality still hinges on the network's ability to disentangle artifacts from true anatomical signals-both of which exhibit complex, multi-scale contextual structure. Convolutional neural networks (CNNs) capture local correlations but often struggle with non-local dependencies. While transformers aim to alleviate this limitation, practical implementations involve design compromises to reduce computational cost by balancing local and non-local sensitivity, occasionally resulting in performance comparable to CNNs. To address these challenges, we propose MambaRoll, a novel physics-driven autoregressive state space model (SSM) for high-fidelity and efficient image reconstruction. MambaRoll employs an unrolled architecture where each cascade autoregressively predicts finer-scale feature maps conditioned on coarser-scale representations, enabling consistent multi-scale context propagation. Each stage is built on a hierarchy of scale-specific PD-SSM modules that capture spatial dependencies while enforcing data consistency through residual correction. To further improve scale-aware learning, we introduce a Deep Multi-Scale Decoding (DMSD) loss, which provides supervision at intermediate spatial scales in alignment with the autoregressive design. Demonstrations on accelerated MRI and sparse-view CT reconstructions show that MambaRoll provides improved performance relative to state-of-the-art CNN-, transformer-, and SSM-based methods.
SN  - 0278-0062
SN  - 1558-254X
ER  -

Abstract

Related

Cite