Quantitative CTA vascular calcification, atherosclerosis burden, and stroke mechanism in patients with ischemic stroke.

BACKGROUND: Vascular calcification is recognized as the advanced stage of atherosclerosis burden. We hypothesized that vascular calcium quantification in CT angiography (CTA) would be helpful to differentiate large artery atherosclerosis (LAA) from other stroke etiology in patients with ischemic stroke. METHODS: We studied 375 acute ischemic stroke patients (200 males, mean age 69.9 years) who underwent complete CTA images of the aortic arch, neck, and head. The automatic artery and calcification segmentation method measured calcification volumes in the intracranial internal carotid artery (ICA), cervical carotid artery, and aortic arch using deep-learning U-net model and region-grow algorithms. We investigated the correlations and patterns of vascular calcification in the different vessel beds among stroke etiology by age category (young: <65 years, intermediate: 65-74 years, older ≥75 years). RESULTS: Ninety-five (25.3%) were diagnosed with LAA according to TOAST criteria. Median calcification volumes were higher by increasing the age category in each vessel bed. One-way ANOVA with Bonferroni correction showed calcification volumes in all vessel beds were significantly higher in LAA compared with other stroke subtypes in the younger subgroup. Calcification volumes were independently associated with LAA in intracranial ICA (OR; 2.89, 95% CI 1.56-5.34, P = .001), cervical carotid artery (OR; 3.40, 95% CI 1.94-5.94, P < .001) and aorta (OR; 1.69, 95%CI 1.01-2.80, P = .044) in younger subsets. By contrast, the intermediate and older subsets did not show a significant relationship between calcification volumes and stroke subtypes. CONCLUSION: Atherosclerosis calcium volumes in major vessels were significantly higher in LAA compared to non-LAA stroke in younger age.

Deep learning-based automatic pipeline for quantitative assessment of thigh muscle morphology and fatty infiltration.

PURPOSE: Fast and accurate thigh muscle segmentation from MRI is important for quantitative assessment of thigh muscle morphology and composition. A novel deep learning (DL) based thigh muscle and surrounding tissues segmentation model was developed for fully automatic and reproducible cross-sectional area (CSA) and fat fraction (FF) quantification and tested in patients at 10 years after anterior cruciate ligament reconstructions. METHODS: A DL model combining UNet and DenseNet was trained and tested using manually segmented thighs from 16 patients (32 legs). Segmentation accuracy was evaluated using Dice similarity coefficients (DSC) and average symmetric surface distance (ASSD). A UNet model was trained for comparison. These segmentations were used to obtain CSA and FF quantification. Reproducibility of CSA and FF quantification was tested with scan and rescan of six healthy subjects. RESULTS: The proposed UNet and DenseNet had high agreement with manual segmentation (DSC >0.97, ASSD < 0.24) and improved performance compared with UNet. For hamstrings of the operated knee, the automated pipeline had largest absolute difference of 6.01% for CSA and 0.47% for FF as compared to manual segmentation. In reproducibility analysis, the average difference (absolute) in CSA quantification between scan and rescan was better for the automatic method as compared with manual segmentation (2.27% vs. 3.34%), whereas the average difference (absolute) in FF quantification were similar. CONCLUSIONS: The proposed method exhibits excellent accuracy and reproducibility in CSA and FF quantification compared with manual segmentation and can be used in large-scale patient studies.

OpenKBP-Opt: an international and reproducible evaluation of 76 knowledge-based planning pipelines.

Objective.To establish an open framework for developing plan optimization models for knowledge-based planning (KBP).Approach.Our framework includes radiotherapy treatment data (i.e. reference plans) for 100 patients with head-and-neck cancer who were treated with intensity-modulated radiotherapy. That data also includes high-quality dose predictions from 19 KBP models that were developed by different research groups using out-of-sample data during the OpenKBP Grand Challenge. The dose predictions were input to four fluence-based dose mimicking models to form 76 unique KBP pipelines that generated 7600 plans (76 pipelines × 100 patients). The predictions and KBP-generated plans were compared to the reference plans via: the dose score, which is the average mean absolute voxel-by-voxel difference in dose; the deviation in dose-volume histogram (DVH) points; and the frequency of clinical planning criteria satisfaction. We also performed a theoretical investigation to justify our dose mimicking models.Main results.The range in rank order correlation of the dose score between predictions and their KBP pipelines was 0.50-0.62, which indicates that the quality of the predictions was generally positively correlated with the quality of the plans. Additionally, compared to the input predictions, the KBP-generated plans performed significantly better (P< 0.05; one-sided Wilcoxon test) on 18 of 23 DVH points. Similarly, each optimization model generated plans that satisfied a higher percentage of criteria than the reference plans, which satisfied 3.5% more criteria than the set of all dose predictions. Lastly, our theoretical investigation demonstrated that the dose mimicking models generated plans that are also optimal for an inverse planning model.Significance.This was the largest international effort to date for evaluating the combination of KBP prediction and optimization models. We found that the best performing models significantly outperformed the reference dose and dose predictions. In the interest of reproducibility, our data and code is freely available.

Automated knee cartilage segmentation for heterogeneous clinical MRI using generative adversarial networks with transfer learning.

Background: This study aimed to build a deep learning model to automatically segment heterogeneous clinical MRI scans by optimizing a pre-trained model built from a homogeneous research dataset with transfer learning. Methods: Conditional generative adversarial networks pretrained on the Osteoarthritis Initiative MR images was transferred to 30 sets of heterogenous MR images collected from clinical routines. Two trained radiologists manually segmented the 30 sets of clinical MR images for model training, validation and test. The model performance was compared to models trained from scratch with different datasets, as well as two radiologists. A 5-fold cross validation was performed. Results: The transfer learning model obtained an overall averaged Dice coefficient of 0.819, an averaged 95 percentile Hausdorff distance of 1.463 mm, and an averaged average symmetric surface distance of 0.350 mm on the 5 random holdout test sets. A 5-fold cross validation had a mean Dice coefficient of 0.801, mean 95 percentile Hausdorff distance of 1.746 mm, and mean average symmetric surface distance of 0.364 mm. It outperformed other models and performed similarly as the radiologists. Conclusions: A transfer learning model was able to automatically segment knee cartilage, with performance comparable to human, using heterogeneous clinical MR images with a small training data size. In addition, the model proved robust when tested through cross validation and on images from a different vendor. We found it feasible to perform fully automated cartilage segmentation of clinical knee MR images, which would facilitate the clinical application of quantitative MRI techniques and other prediction models for improved patient treatment planning.

Automatic segmentation of gadolinium-enhancing lesions in multiple sclerosis using deep learning from clinical MRI.

Gadolinium-enhancing lesions reflect active disease and are critical for in-patient monitoring in multiple sclerosis (MS). In this work, we have developed the first fully automated method to segment and count the gadolinium-enhancing lesions from routine clinical MRI of MS patients. The proposed method first segments the potential lesions using 2D-UNet from multi-channel scans (T1 post-contrast, T1 pre-contrast, FLAIR, T2, and proton-density) and classifies the lesions using a random forest classifier. The algorithm was trained and validated on 600 MRIs with manual segmentation. We compared the effect of loss functions (Dice, cross entropy, and bootstrapping cross entropy) and number of input contrasts. We compared the lesion counts with those by radiologists using 2,846 images. Dice, lesion-wise sensitivity, and false discovery rate with full 5 contrasts were 0.698, 0.844, and 0.307, which improved to 0.767, 0.969, and 0.00 in large lesions (>100 voxels). The model using bootstrapping loss function provided a statistically significant increase of 7.1% in sensitivity and of 2.3% in Dice compared with the model using cross entropy loss. T1 post/pre-contrast and FLAIR were the most important contrasts. For large lesions, the 2D-UNet model trained using T1 pre-contrast, FLAIR, T2, PD had a lesion-wise sensitivity of 0.688 and false discovery rate 0.083, even without T1 post-contrast. For counting lesions in 2846 routine MRI images, the model with 2D-UNet and random forest, which was trained with bootstrapping cross entropy, achieved accuracy of 87.7% using T1 pre-contrast, T1 post-contrast, and FLAIR when lesion counts were categorized as 0, 1, and 2 or more. The model performs well in routine non-standardized MRI datasets, allows large-scale analysis of clinical datasets, and may have clinical applications.

The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset.

PURPOSE: To organize a multi-institute knee MRI segmentation challenge for characterizing the semantic and clinical efficacy of automatic segmentation methods relevant for monitoring osteoarthritis progression. MATERIALS AND METHODS: A dataset partition consisting of three-dimensional knee MRI from 88 retrospective patients at two time points (baseline and 1-year follow-up) with ground truth articular (femoral, tibial, and patellar) cartilage and meniscus segmentations was standardized. Challenge submissions and a majority-vote ensemble were evaluated against ground truth segmentations using Dice score, average symmetric surface distance, volumetric overlap error, and coefficient of variation on a holdout test set. Similarities in automated segmentations were measured using pairwise Dice coefficient correlations. Articular cartilage thickness was computed longitudinally and with scans. Correlation between thickness error and segmentation metrics was measured using the Pearson correlation coefficient. Two empirical upper bounds for ensemble performance were computed using combinations of model outputs that consolidated true positives and true negatives. RESULTS: Six teams (T 1-T 6) submitted entries for the challenge. No differences were observed across any segmentation metrics for any tissues (P = .99) among the four top-performing networks (T 2, T 3, T 4, T 6). Dice coefficient correlations between network pairs were high (> 0.85). Per-scan thickness errors were negligible among networks T 1-T 4 (P = .99), and longitudinal changes showed minimal bias (< 0.03 mm). Low correlations (ρ < 0.41) were observed between segmentation metrics and thickness error. The majority-vote ensemble was comparable to top-performing networks (P = .99). Empirical upper-bound performances were similar for both combinations (P = .99). CONCLUSION: Diverse networks learned to segment the knee similarly, where high segmentation accuracy did not correlate with cartilage thickness accuracy and voting ensembles did not exceed individual network performance.See also the commentary by Elhalawani and Mak in this issue.Keywords: Cartilage, Knee, MR-Imaging, Segmentation © RSNA, 2020Supplemental material is available for this article.

Automated cartilage and meniscus segmentation of knee MRI with conditional generative adversarial networks.

PURPOSE: Fully automatic tissue segmentation is an essential step to translate quantitative MRI techniques to clinical setting. The goal of this study was to develop a novel approach based on the generative adversarial networks for fully automatic segmentation of knee cartilage and meniscus. THEORY AND METHODS: Defining proper loss function for semantic segmentation to enforce the learning of multiscale spatial constraints in an end-to-end training process is an open problem. In this work, we have used the conditional generative adversarial networks to improve segmentation performance of convolutional neural network, such as UNet alone by overcoming the problems caused by pixel-wise mapping based objective functions, and to capture cartilage features during the training of the network. Furthermore, the Dice coefficient and cross entropy losses were incorporated to the loss functions to improve the model performance. The model was trained and tested on 176, 3D DESS (double-echo steady-state) knee images from the Osteoarthritis Initiative data set. RESULTS: The proposed model provided excellent segmentation performance for cartilages with Dice coefficients ranging from 0.84 in patellar cartilage to 0.91 in lateral tibial cartilage, with an average Dice coefficient of 0.88. For meniscus segmentation, the model achieves 0.89 Dice coefficient for lateral meniscus and 0.87 Dice coefficient for medial meniscus. The results are superior to previously published automatic cartilage and meniscus segmentation methods based on deep learning models such as convolutional neural network. CONCLUSION: The proposed UNet-conditional generative adversarial networks based model demonstrated a fully automated segmentation method with high accuracy for knee cartilage and meniscus.