Automatic segmentation of the aorta on multi-center and multi-vendor phase-contrast enhanced magnetic resonance angiographies and the advantages of transfer learning

Funding Acknowledgements

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Guala A. received funding from the Spanish Ministry of Science, Innovation and Universities

Background

Phase-contrast (PC) enhanced magnetic resonance (MR) angiography (MRA) is a class of angiogram exploiting velocity data to increase the signal-to-noise ratio, thus avoiding the administration of external contrast agent, normally used to segment 4D flow MR data. To train deep-learning algorithms to segment PC-MRA a large amount of manually annotated data is needed: however, the relatively novelty of the sequence, its rapid evolution and the extensive time needed to manually segment data limit its availability.

Purpose

The aim of this study was to test a deep learning algorithm in the segmentation of multi-center and multi-vendor PC-MRA and to test if transfer learning (TL) improves performance.

Methods

A large dataset (LD) of 262 and a small one (SD) of 22 PC-MRA, acquired without contrast agent at 1.5 T in a General Electric and a Siemens scanner, respectively, were manually annotated and divided into training (232 and 15 cases) and testing (30 and 7) sets. They both included PC-MRA of healthy subjects and patients with aortic diseases (excluding dissections) and native aorta. A convolutional neural networks (CNN) based on nnU-Net framework [1] was trained in the LD and another in the SD. The left ventricle was removed semi-automatically from the DL segmentations of the LD as it was not relevant for this application. Networks were then tested on the test sets of the dataset there were trained and the other dataset to assess generalizability. Finally, a fine-tuning transfer learning approach was applied to LD network and the performance on both test sets were tested. Dice score, Hausdorff distance, Jaccard score and Average Symmetrical Surface Distance were used as segmentation quality metrics.

Results

LD network achieved good performance in LD test set, with a DS of 0.904, ASSD of 1.47, J of 0.827 and HD of 6.35, which further improve after removing the left ventricle in the post-processing to a DS of 0.942, ASSD of 0.93, J of 0.892 and HD of 3.32. SD network results in an average DS of 0.895, ASSD of 0.59, J of 0.812 and HD of 2.05. Once tested on the testing set of the other dataset, LD network resulted in a DS of 0.612 while SD network in DS of 0.375, thus showing limited generalizability. However, the application of transfer learning to LD network resulted in the improvement of the evaluation metrics on the SD from a DS of 0.612 to 0.858, while slightly worsening in the first one without post-processing to 0.882.

Conclusions

nnU-net framework is effective for fast automatic segmentation of the aorta from multi-center and multi-vendor PC-MRA, showing performance comparable with the state of the art. The application of transfer learning allows for increased generalization to data from center not included in the original training. These results unlock the possibility for fully-automatic analysis of multi-vendor multi-center 4D flow MR.

Machine learning to automatically detect anatomical landmarks on phase-contrast enhanced magnetic resonance angiography

Funding Acknowledgements

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Guala A. received funding from the Spanish Ministry of Science, Innovation and Universities

Introduction

The heterogeneous characteristic of the thoracic aorta implies that all biomarkers with potential for risk stratification need to be references to a specific location. This is the case, for example, of diameter [1], stiffness [2] and wall shear stress [3]. This is normally achieved by the manual identification of a limited number of key anatomic landmarks [4], which is a time-demanding task and may impact biomarkers accuracy and reproducibility. Automatic identification of these anatomic landmarks may speed-up the analysis and allow for the creation of fully automatic image analysis pipelines. Machine learning (ML) algorithms might be suitable for this task.

Purpose

The aim of this study was to test the performance of a ML algorithm in localizing key thoracic anatomical landmarks on phase-contrast enhanced magnetic resonance angiograms (PC-MRA).

Methods

PC-MRA of 323 patients with native aorta and aortic valve and a variety of aortic conditions (141 bicuspid aortic valve patients, 60 patients with degenerative aortic aneurysms, 82 patients with genetic aortopathy and 40 healthy volunteers) were included in this study. Four anatomical landmarks were manually identified on PC-MRA by an experienced researcher: sinotubular junction, the pulmonary artery bifurcation and the first and third supra-aortic vessel braches. A reinforcement learning algorithm (DQN), combining Q-learning with deep neural networks, was trained. The algorithm was tested in a separate set of 30 PC-MRA with similar distribution of aortic conditions in which human intra-observer reproducibility was quantified. The distance between points was used as quality metric and human annotation was considered as ground-truth. Repeated-measures ANOVA was used for statistical testing.

Results

ML algorithm resulted in performance similar to the intra-observer variability obtained by the experienced human reader in the identification of the sinotubular junction (11.1 ± 8.6 vs 11.0 ± 8.1 mm, p = 0.949) and first (6.8 ± 5.6 vs 6.6 ± 3.9 mm, p = 0.886) and third (8.4 ± 7.4 vs 6.8 ± 4.0 mm, p = 0.161) supra-aortic vessels branches. However, the algorithm did not reach human-level performance in the localization of the pulmonary artery bifurcation (15.2 ± 13.1 vs 10.2 ± 7.0 mm, p = 0.008). The time needed to the ML algorithm to locate all points ranged between 0.8 and 1.6 seconds on a standard computer while manual annotation required around two minutes to be performed.

Conclusions

The rapid identification of key aortic anatomical landmarks by a reinforced learning algorithm is feasible with human-level performance. This approach may thus be used for the design of fully-automatic pipeline for 4D flow CMR analysis.

Reinforcement machine learning-based aortic anatomical landmarks detection from phase-contrast enhanced magnetic resonance angiography

Funding Acknowledgements

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Spanish Ministry of Science, Innovation and Universities; La Marató de TV3

Introduction

Automatic analysis of medical imaging data may improve their clinical impact by reducing analysis time and improving reproducibility. Many medical imaging data, like 4D-flow magnetic resonance imaging (MRI), are often quantified regionally, implying the need for anatomical landmark identification to locate correspondences in the extracted data and compare among patients. Machine learning (ML) techniques hold potential for automatic analysis of medical imaging. Phase-contrast enhanced magnetic resonance angiography (PC-MRA) is a class of angiograms not requiring the administration of contrast agents.

Purpose

We aimed to test whether a machine learning algorithm can be trained to identify key anatomical cardiovascular landmarks on PC-MRA images and compare its performance with humans.

Methods

Three-hundred twenty-three aortic PC-MRA were manually annotated with the location of 4 landmarks: sinotubular junction, pulmonary artery bifurcation and first and third supra-aortic vessels (Figure 1), often used to separate the aorta in sub-regions. Patients included in the training dataset comprised healthy volunteers (40), bicuspid aortic valve patients (141), patients with degenerative aortic disease (60) and patients with genetically-triggered aortic disease (82), all without previous aortic surgery and with native aortic valve. PC-MRA images and manual annotations were used to train a DQN, a reinforcement learning algorithm that combines Q-learning with deep neural networks. The agents can navigate the images and optimally find the landmarks by following the policies learned during training. Data from thirty patients, distributed in terms of aortic condition as the training set, unseen by the algorithm in the training phase, were used to quantify intra-observer reproducibility and to assess ML algorithm performance. Distance between points was used as metric for comparisons, original human annotation was used as ground-truth and repeated-measures ANOVA was used for statistical testing.

Results

Human and machine learning performed similarly in the identification of the sinotubular junction (distance between points of 11.0 ± 8.1 vs. 11.1 ± 8.6 mm, respectively, p = 0.949) and first (6.6 ± 3.9 vs. 6.8 ± 5.6 mm, p = 0.886) and third (6.8 ± 4.0 vs. 8.4 ± 7.4 mm, p = 0.161) supra-aortic vessels branches but human annotation outperformed ML landmark detection in the identification of the pulmonary artery bifurcation (10.2 ± 7.0 vs. 15.2 ± 13.1 mm, p = 0.008). Computation time for landmark detection by ML was between 0.8 and 1.6 seconds on a standard computer while human annotation took approximatively two minutes.

Conclusions

ML-based aortic landmarks detection from phase-contrast enhanced magnetic resonance angiography is feasible and fast and performs similarly to human. Reinforced learning anatomical landmark identification unlock automatic extraction of a variety of regional aortic data, including complex 4D flow parameters.

Abstract Figure