Assessment of Robustness of MRI Radiomic Features in Four Abdominal Organs: Impact of Deep Learning Reconstruction and Segmentation.
Authors
Affiliations (11)
Affiliations (11)
- Department of Imaging, Tongren Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Shanghai Key Laboratory of Flexible Medical Robotics, Tongren Hospital, Institute of Medical Robotics, Shanghai Jiao Tong University, Shanghai, China.
- Department of Epidemiology and Population Health, Stanford University School of Medicine, Stanford, California, USA.
- Department of Biomedical Engineering, Boston University, Boston, Massachusetts, USA.
- Jacobi Medical Center, Albert Einstein College of Medicine, New York, New York, USA.
- MR Scientific Marketing, Siemens Healthineers Ltd., Shanghai, China.
- MR Application, Siemens Healthineers Ltd., Shanghai, China.
- MR Application Predevelopment, Siemens Healthcare, Erlangen, Germany.
- Department of Radiology, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- College of Health Science and Technology, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
- Shanghai Key Laboratory of Gastric Neoplasms, Department of Surgery, Shanghai Institute of Digestive Surgery, Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China.
Abstract
The impact of deep learning (DL) reconstruction and segmentation on MRI radiomics stability has not been fully assessed. To investigate the effects of acquisition, reconstruction, and segmentation on the reproducibility and variability of radiomic features in abdominal MRI. Prospective. 37 volunteers (22 men; mean age ± standard deviation, 37.4 ± 11.0 years). 3.0-T; axial turbo spin echo T2-weighted image, and fat-suppressed T2-weighted image using a half-Fourier acquisition single-shot turbo spin echo technique, each acquired four times with conventional or accelerated techniques, reconstructed with standard or DL algorithms. Regions of interest were automatically generated by a DL neural network for liver, spleen, and right and left kidneys, followed by manual correction. We extracted 107 features using PyRadiomics after z-score normalization. The reproducibility between acquisitions, reconstructions, and segmentations was evaluated using intraclass correlation coefficient (ICC) and concordance correlation coefficient (CCC). The variability among the four scans was assessed by coefficient of variation (CV) and quartile coefficient of dispersion (QCD). p < 0.05 was considered significant. The mean ICC (0.518-0.608; 0.606-0.681) and CCC (0.515-0.603; 0.601-0.680) values were low for both manual and automatic segmentation regardless of image acquisition and reconstruction, using conventional acquisition with standard reconstruction as reference. The mean ICC (0.535-0.713) and CCC (0.531-0.714) values were low between manual and automatic segmentation, regardless of image acquisition and reconstruction. The median CV (10.0%-17.5%; 8.9%-15.5%) and QCD (5.3%-8.5%; 5.1%-8.3%) values were moderate but still adequate for both manual and automatic segmentation among different scans. Given the substantial impact of accelerated acquisition and DL reconstruction on the robustness of radiomics features in abdominal MRI, caution should be exercised when utilizing images with different acquisition and reconstruction techniques in radiomics analysis. The automatic segmentation cannot replace manual segmentation due to insufficient robustness of radiomics features. 2. Stage 1.