Overview

Schema Version

https://atlas.rsna.org/schemas/2025-11/model.json

Name

Masked-LMCTrans (Longitudinal Multimodality Coattentional CNN Transformer)

Link

https://dx.doi.org/10.1148/ryai.220246

Indexing

Keywords: Pediatrics, PET, PET/MRI, FDG, Lymphoma, Dose reduction, Ultra-low-dose, Transformer, Coattention, Longitudinal reconstruction, CNN, DenseNet

Content: MI, NM, PD, MR

RadLex: RID12963, RID11701, RID16802, RID12688, RID10337

SNOMED: 1172592001, 118601006, 1163005009

Author(s)

Yan-Ran (Joyce) Wang

Liangqiong Qu

Natasha Diba Sheybani

Xiaolong Luo

Jiangshan Wang

Kristina Elizabeth Hawk

Ashok Joseph Theruvath

Sergios Gatidis

Xuerong Xiao

Allison Pribnow

Daniel Rubin

Heike E. Daldrup-Link

Organization(s)

Stanford University, Departments of Biomedical Data Science, Radiology, and Nuclear Medicine

University Hospital Tübingen, Department of Diagnostic and Interventional Radiology

University of Science and Technology of China, School of Engineering

Lucile Packard Children's Hospital, Stanford University School of Medicine, Department of Pediatrics, Division of Pediatric Oncology

Version

1.0

License

Contact

Yan-Ran (Joyce) Wang, email: moc.liamg@001narnaygnaw

Funding

Supported by the National Cancer Institute of the U.S. National Institutes of Health (grant R01CA269231) and the Andrew McDonough B+ Foundation.

Ethical review

HIPAA-compliant retrospective study; IRB approvals obtained at participating centers; written informed consent obtained from adult patients and parents/guardians; pediatric assent obtained.

Date

Published: 2023-05-03

Created: 2022-11-10

References

[1] Wang YR, Qu L, Sheybani ND, Luo X, Wang J, Hawk KE, Theruvath AJ, Gatidis S, Xiao X, Pribnow A, Rubin D, Daldrup-Link HE. "AI Transformers for Radiation Dose Reduction in Serial Whole-Body PET Scans". Radiology: Artificial Intelligence. 2023 May;5(3):e220246.. 2023-05-03. doi:10.1148/ryai.220246. PMID: 37293349. PMCID: PMC10245181.

Model

Architecture

Longitudinal multimodality coattentional CNN transformer with separate PET and MRI DenseNet encoders and coattentional transformer layers enabling cross-attention between baseline and follow-up streams; 2D slice-wise inference with 3D volume reconstruction.

Availability

All code, trained models, and de-identified data are available from the authors upon reasonable request.

Clinical benefit

Enables reconstruction of high-quality follow-up PET images from 1% of standard counts, supporting substantial radiotracer dose reduction and/or faster scans, potentially improving safety and throughput for pediatric oncology imaging.

Clinical workflow phase

Post-acquisition image reconstruction/post-processing.

Degree of automation

Automated image reconstruction algorithm; assists clinicians by producing enhanced PET images from ultra-low-dose inputs.

Indications for use

Reconstruction of follow-up whole-body PET images in patients undergoing serial PET/MRI, demonstrated in pediatric patients with lymphoma in a retrospective study; imaging department/clinical radiology environment.

Input

Four inputs: baseline PET (with tumor regions masked), baseline MRI, follow-up 1% count PET, and follow-up MRI; processed as 2D sections for whole-body reconstruction.

Instructions

Provide baseline PET with prominent tumor regions masked and baseline MRI as references, along with follow-up 1% PET/MRI. The method performs 2D slice-wise inference and stacks slices to form 3D output. Implemented in Python 3 with PyTorch 1.10.

Limitations

Evaluated retrospectively on pediatric lymphoma; simulated ultra-low-dose PET (1% counts) rather than prospectively injected low-dose; performance in subtle/low-grade lesions, bone, and new lesions not established; limited dataset size; hardware/software variations exist though external cohort showed generalizability; evolving PET detector technologies may impact comparative performance.

Output

CDEs: RDE2721, RDE1872

Description: Reconstructed follow-up whole-body PET image approximating standard-dose image quality from 1% count input using longitudinal PET/MRI references.

Recommendation

Research use with further prospective validation required before clinical deployment, including reader studies and evaluation in true injected ultra-low-dose scenarios.

Regulatory information

Authorization status: Research study; no regulatory clearance reported.

Reproducibility

Implemented in Python 3 using PyTorch 1.10; performance metrics computed in Python; statistical analyses in R. Code/models/data available upon reasonable request.

Use

Intended: Image reconstruction

Out-of-scope: Detection and diagnosis

Excluded: Image reconstruction

User

Intended: Radiologist, Subspecialist diagnostic radiologist, Researcher

Out-of-scope: Patient

Excluded: Layperson