Cardiac magnetic resonance imaging (CMR) is a widely used non-invasive imaging modality for evaluating cardiovascular diseases. CMR is the gold standard method for left and right ventricular functional assessment due to its ability to characterize myocardial structure and function and low intra- and inter-observer variability. However the post-processing segmentation during the functional evaluation is time-consuming and challenging. A fully automated segmentation method can assist the experts; therefore, they can do more efficient work. In this paper, a regression-based fully automated method is presented for the right- and left ventricle segmentation. For training and evaluation, our dataset contained MRI short-axis scans of 5570 patients, who underwent CMR examinations at Heart and Vascular Center, Semmelweis University Budapest. Our approach is novel and after training the state-of-the-art algorithm on our dataset, our algorithm proved to be superior on both of the ventricles. The evaluation metrics were the Dice index, Hausdorff distance and volume related parameters. We have achieved average Dice index for the left endocardium: 0.927, left epicardium: 0.940 and right endocardium: 0.873 on our dataset. We have also compared the performance of the algorithm to the human-level segmentation on both ventricles and it is similar to experienced readers for the left, and comparable for the right ventricle. We also evaluated the proposed algorithm on the ACDC dataset, which is publicly available, with and without transfer learning. The results on ACDC were also satisfying and similar to human observers. Our method is lightweight, fast to train and does not require more than 2 GB GPU memory for execution and training.

Left ventricular hypertrophy is an independent predictor of coronary artery disease, stroke, and heart failure. Our aim was to detect LVH cardiac magnetic resonance (CMR) scans with automatic methods. We developed an ensemble model based on a three-dimensional version of ResNet. The input of the network included short-axis and long-axis images. We also introduced a standardization methodology to unify the input images for noise reduction. The output of the network is the decision whether the patient has hypertrophy or not. We included 428 patients (mean age: 49 ± 18 years, 262 males) with LVH (346 hypertrophic cardiomyopathy, 45 cardiac amyloidosis, 11 Anderson–Fabry disease, 16 endomyocardial fibrosis, 10 aortic stenosis). Our control group consisted of 234 healthy subjects (mean age: 35 ± 15 years; 126 males) without any known cardiovascular diseases. The developed machine-learning-based model achieved a 92% F1-score and 97% recall on the hold-out dataset, which is comparable to the medical experts. Experiments showed that the standardization method was able to significantly boost the performance of the algorithm. The algorithm could improve the diagnostic accuracy, and it could open a new door to AI applications in CMR.

Current video streaming services use a conventional, client-server network topology that puts a heavy load on content servers. Previous work has shown that Peer-to-Peer (P2P) assisted streaming solutions can potentially reduce this load. However, implementing P2P-assisted streaming poses several challenges in modern networks. Users tend to stream videos on the go, using their mobile devices. This mobility makes the network difficult to orchestrate. Furthermore, peers have to contribute their storage to the network, which is challenging, since mobile devices have limited resources compared to desktop machines. In this paper, we introduce an analytical framework for mobile P2P-assisted streaming to estimate the server load that we define as the minimum required server upload rate. Using our framework, we evaluate four caching strategies: infinite cache as a baseline, first in first out (FIFO), random, and Random Linear Network Coded (RLNC) cache. We verify our anal tical results with empirical data that was obtained by carrying out extensive measurements on our working P2P system. Our results show that when employing FIFO, random, and RLNC caching strategies, the server load converges to that of the infinite cache as the cache size increases. With a limit of 5 P2P connections per peer, we show that using the random caching, peers can store 40% fewer packets and still achieve the same benefit as with FIFO caching. When using the RLNC caching, it is enough to store 50% fewer packets to achieve the same benefit.