Zero echo time MRI with deep learning reconstruction and chemical shift correction for detecting osteolytic myeloma lesions.
Authors
Affiliations (7)
Affiliations (7)
- Department of Medical Imaging, Institut de Recherche Expérimentale et Clinique (IREC), Institut du Cancer Roi Albert II, Cliniques Universitaires Saint Luc, Université catholique de Louvain (UCLouvain), Brussels, Belgium.
- GE HealthCare MR Clinical Solutions, Atlanta, GA, USA.
- GE HealthCare MR Clinical Solutions, New York, NY, USA.
- GE HealthCare, Buc, France.
- Department of Hematology, Institut de Recherche Expérimentale et Clinique (IREC), Institut du Cancer Roi Albert II, Cliniques Universitaires Saint Luc, Université catholique de Louvain (UCLouvain), Brussels, Belgium.
- Department of Nuclear Medicine, Institut de Recherche Expérimentale et Clinique (IREC), Institut du Cancer Roi Albert II, Cliniques Universitaires Saint Luc, Université catholique de Louvain (UCLouvain), Brussels, Belgium.
- Department of Medical Imaging, Institut de Recherche Expérimentale et Clinique (IREC), Institut du Cancer Roi Albert II, Cliniques Universitaires Saint Luc, Université catholique de Louvain (UCLouvain), Brussels, Belgium. [email protected].
Abstract
We compared three magnetic resonance imaging (MRI) sequences-native zero echo time (ZTE), deep learning (DL)-chemical shift correction (CSC) reconstructed ZTE (ZTE-DLCSC), and gradient-echo black bone (BB)-in detecting osteolytic multiple myeloma (MM) lesions, using computed tomography (CT) as reference. Patients who had undergone whole-body <sup>18</sup>F-fluorodeoxiglucose-positron emission tomography/CT were prospectively enrolled at a single-center and underwent 3-T whole-body MRI with ZTE and BB sequences covering the lumbar spine, pelvis, and proximal femurs. ZTE-DLCSC images were reconstructed from raw data. Ten bone regions were assessed for lesion presence and number by a senior radiologist with 26 years' experience and two junior readers (a radiology fellow and a resident). Repeatability and reproducibility (Gwet's agreement coefficients AC1 and AC2), differences in quantitative counts, and accuracy were evaluated per-sequence/region/reader. Ten patients, aged 67 ± 12 years (mean ± standard deviation), were enrolled. Considering all regions, repeatability was at least moderate for ZTE (AC1 ≥ 0.45), good for ZTE-DLCSC and BB (AC1 ≥ 0.60) and very good for CT (AC1 ≥ 0.80). Reproducibility was at least fair for ZTE and BB (AC2 ≥ 0.20), good for ZTE-DLCSC (AC2 ≥ 0.60) and very good for CT (AC2 ≥ 0.80). Accuracy of ZTE-DLCSC ranged 80‒93%; compared to ZTE, accuracy increased by 23% and 25% for junior readers (p = 0.010 and p = 0.002 respectively) and by 32% for the senior reader (p < 0.001). ZTE-DLCSC detected more lesions than ZTE/BB (+30%, p = 0.011; +25%, p = 0.024). DL-driven DLCSC reconstruction improves the reliability and accuracy of the ZTE sequence for detecting MM lesions. The combination of deep learning reconstruction and chemical shift correction improves the accuracy of zero echo time MRI for detecting myeloma lesions. This study was approved by the institutional ethics committee and registered at ClinicalTrials.gov (NCT05381077). The performance of ZTE-DLCSC for detecting osteolytic MM lesions was compared with that of native ZTE and BB sequences using CT as reference. ZTE-DLCSC showed better intra-reader agreement (repeatability) and inter-reader agreement (reproducibility) than ZTE and BB sequences. ZTE-DLCSC had fewer false positives and false negatives than ZTE and BB sequences. ZTE-DLCSC detected more lesions than ZTE and BB sequences.