Amir Averbuch

Brugada syndrome (BrS) is an inherited electrical cardiac disorder that is associated with a higher risk of ventricular fibrillation (VF) and sudden cardiac death (SCD) in patients without structural heart disease. The diagnosis is based on the documentation of the typical pattern in the electrocardiogram (ECG) characterized by a J-point elevation of ≥2 mm, coved-type ST-segment elevation, and negative T wave in one or more right precordial leads, called type 1 Brugada ECG. Risk stratification is particularly difficult in asymptomatic cases. Patients who have experienced documented VF are generally recommended to receive an implantable cardioverter defibrillator to lower the likelihood of sudden death due to recurrent episodes. However, for asymptomatic individuals, the most appropriate course of action remains uncertain. Accurate risk prediction is critical to avoiding premature deaths and unnecessary treatments. Due to the challenges associated with experimental research on human cardiac tissue, alternative techniques such as computational modeling and deep learning-based artificial intelligence (AI) are becoming increasingly important. This study introduces a vision transformer (ViT) model that leverages 12-lead ECG images to predict potentially fatal arrhythmic events in BrS patients. This dataset includes a total of 278 ECGs, belonging to 210 patients which have been diagnosed with Brugada syndrome, and it is split into two classes: event and no event. The event class contains 94 ECGs of patients with documented ventricular tachycardia, ventricular fibrillation, or sudden cardiac death, while the no event class is composed of 184 ECGs used as the control group. At first, the ViT is trained on a balanced dataset, achieving satisfactory results (89% accuracy, 94% specificity, 84% sensitivity, and 89% F1-score). Then, the discarded no event ECGs are attached to additional 30 event ECGs, extracted by a 24 h recording of a singular individual, composing a new test set. Finally, the use of an optimized classification threshold improves the predictions on an unbalanced set of data (74% accuracy, 95% negative predictive value, and 90% sensitivity), suggesting that the ECG signal can reveal key information for the risk stratification of patients with Brugada syndrome.

Modern cameras typically offer two types of image states: a minimally processed linear raw RGB image representing the raw sensor data, and a highly-processed non-linear image state, such as the sRGB state. The CIE-XYZ color space is a device-independent linear space used as part of the camera pipeline and can be helpful for computer vision tasks, such as image deblurring, dehazing, and color recognition tasks in medical applications, where color accuracy is important. However, images are usually saved in non-linear states, and achieving CIE-XYZ color images using conventional methods is not always possible. To tackle this issue, classical methodologies have been developed that focus on reversing the acquisition pipeline. More recently, supervised learning has been employed, using paired CIE-XYZ and sRGB representations of identical images. However, obtaining a large-scale dataset of CIE-XYZ and sRGB pairs can be challenging. To overcome this limitation and mitigate the reliance on large amounts of paired data, self-supervised learning (SSL) can be utilized as a substitute for relying solely on paired data. This paper proposes a framework for using SSL methods alongside paired data to reconstruct CIE-XYZ images and re-render sRGB images, outperforming existing approaches. The proposed framework is applied to the sRGB2XYZ dataset.

Learning from imbalanced data is one of the most significant challenges in real-world classification tasks. In such cases, neural networks performance is substantially impaired due to preference towards the majority class. Existing approaches attempt to eliminate the bias through data re-sampling or re-weighting the loss in the learning process. Still, these methods tend to overfit the minority samples and perform poorly when the structure of the minority class is highly irregular. Here, we propose a novel deep meta-learning technique to augment a given imbalanced dataset with new minority instances. These additional data are incorporated in the classifier's deep-learning process, and their contributions are learned explicitly. The advantage of the proposed method is demonstrated on synthetic and real-world datasets with various imbalance ratios.

Deep Learning frameworks are gaining increased popularity in image processing tasks such as computational hyperspectral imaging. While these frameworks achieve state-of-the-art results in terms of reconstruction quality and run time, they often require massive databases of hyperspectral cubes for training the reconstruction algorithms. Unfortunately, such databases are usually hard to acquire due to complexity and cost considerations. To mitigate these challenges, we propose a method for generating a synthetic database of hyperspectral cubes in the visible range using a limited number of natural hyperspectral cubes, an unlimited number of RGB images, and a Generative Adversarial Network model. The suggested algorithm, dubbed SHS-GAN, is trained to get a query RGB image and to output a synthetic hyperspectral cube. While the spectral domain of the synthetic hyperspectral cube shares similar statistical properties as the natural hyperspectral cubes used in the training process, the SHS-GAN is trained to preserve the spatial characteristics of the query RGB image, whereas the R, G, B values provide an additional constraint along with the spectral domain. Our suggested framework was utilized for performing Snapshot Spectral Imaging (SSI) from a single monochromatic dispersed and diffused snapshot using the DD-Net reconstruction neural network. We demonstrate, by simulations and lab experiments, that enhancing the training database with synthetic data from the SHS-GAN improves the reconstruction quality of the hyperspectral cube. In addition, we share a new original database of more than 10,000 hyperspectral cubes of real objects of size 256x256x29 in the 420-700 nm visible range.