KEYWORDS: Kidney, Image segmentation, Magnetic resonance imaging, 3D modeling, Data modeling, 3D image processing, Tumor growth modeling, Algorithm development, Tissues, Cancer
Multi-parametric magnetic resonance imaging (mp-MRI) is a promising tool for diagnosis of renal masses and may outperform computed tomography (CT) to differentiate between benign and malignant renal masses due to superior soft tissue contrast. Deep learning (DL)-based methods for kidney segmentation are under-explored in mp-MRI which consists of several pulse sequences, including primarily T2-weighted (T2W) and contrast-enhanced (CE) images. Multi-parametric MRI images have domain shift due to differences in acquisition systems and image protocols, leading to lack of generalizability of methods for image segmentation. To perform similar automated kidney segmentation on another mp- MRI sequence, the model needs a large dataset with manual segmentations to train a model from scratch, which is labor intensive and time consuming. In this paper, we first trained a DL-based method using 108 cases of labeled data to contour kidneys using T1 weighted-Nephrographic Phase CE-MRI (T1W-NG). We then applied a transfer learning approach to other mp-MRI images using pre-trained weights from the source domain, thus eliminating the need for large manually annotated datasets in target domain. The fully automated 2D U-Net for kidney segmentation in source domain containing total 108 3D images of T1W-NG, yielded Dice-similarity coefficient (DSC) of 0.91 ± 0.07 on test cases. The transfer learning of pretrained weights of T1W-NG model on the smaller target domain T2W dataset containing total 50 3D images for automated kidney segmentation generated DSC of 0.90 ± 0.06 (p<0.05), which was an improvement of 3.43% in DSC by compared to the without transfer learning approach (T2W-UNet model).
KEYWORDS: Image segmentation, Cardiovascular magnetic resonance imaging, Convolutional neural networks, Magnetic resonance imaging, Image processing algorithms and systems, Analytical research
Extra-cellular volume (ECV) mapping cardiac magnetic resonance (CMR) imaging allows for the characterization of expanded myocardial extracellular space, a common feature of myocardial fibrosis (MF). Quantification of MF is feasible using ECV mapping techniques; however, prior manual delineation of the endocardial and epicardial borders is required. In this study, we propose a method for automated myocardial delineation of ECV maps using convolutional neural networks (CNNs). We compare two methods based on the standard U-Net and the U-Net++ architectures using a five-fold cross validation on basal, mid, and apical short-axis ECV maps of the left ventricle (LV) in 73 patients with ischemic (n=38) or dilated (n=35) cardiomyopathies. The standard U-Net and U-Net++ -based architectures yielded DSC metrics of 87.61% and 87.89%, respectively, against manual contours derived by an expert. Precision and recall were reported >85% and relative error <12% for both CNNs. The U-Net++ architecture outperformed the standard U-Net on the order of 1-2% for all metrics. An inter-operator variability analysis was performed on a subset of myocardial contours derived by three operators. The inter-operator analysis demonstrated significant differences in the distribution of myocardial ECV values among three operators as per the Kruskal-Wallis H-test (average pair-wise P-value = 0.040), but operator differences failed to show significance against U-Net++ or standard U-Net (average pair-wise P-value 0.055 and 0.060, respectively). Correlation of global ECV improved for operators against U-Net++ (𝜌=0.88) and against standard U-Net (𝜌=0.877) compared to correlation of global ECV values between all operators (ρ=0.828).
T1-mapping cardiac magnetic resonance (CMR) is a rapidly expanding non-invasive tool for quantitative assessment of myocardial fibrosis. To achieve both efficiency and reproducibility in quantification of T1 measures, automated myocardial boundary tracing is desirable. Accordingly, the application of robust segmentation algorithms for this modality are of significant interest. However, conventional algorithms may fail in myocardial segmentation of T1-mapping images due to low signal gradients at the endocardial-blood pool boundary. In this work, we propose using prior information from cinematic (CINE) CMR images toward accurate myocardial segmentation of native T1-mapping images, acquired using the shortened modified Look-Locker imaging (shMOLLI) technique. We use a three-step framework, which begins with pre-processing and resizing of both CINE and shMOLLI images. Next, we implement semi-automated segmentation of the myocardium on resized CINE images using a deformable model-based technique, via the freely available software Segment v2.2. The final step of our framework is registration and propagation of the CINE contours to corresponding (slice-matched) native shMOLLI images using a non-rigid registration technique based on a modality independent neighborhood descriptor (MIND). We validate our technique on 20 image sets obtained from 20 patients with confirmed myocardial fibrosis related to ischemic injury (myocardial infarction). Our method achieved an average Dice similarity coefficient (DSC) of 84.36% ± 4.03%, precision of 91.68% ± 7.89%, recall of 91.33% ± 8.41% and relative area error of 16.29% ± 8.58%.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.