Unsupervised mitral valve segmentation in echocardiography with neural network matrix factorization.

Abstract

Mitral valve segmentation is a crucial first step to establish a machine learning pipeline that can support practitioners into performing the diagnosis of mitral valve diseases, surgical planning, and intraoperative procedures. To this end, we propose a totally automated and unsupervised mitral valve segmentation algorithm, based on a neural network low-dimension matrix factorization of the echocardiography video. The method is evaluated in a collection of echocardiography video of patients with a variety of mitral valve diseases and exceeds the state-of-the-art method in all the metrics considered.

Neural collaborative filtering for unsupervised mitral valve segmentation in echocardiography

“…In this paper, we propose NN-MitralSeg, an unsupervised MV segmentation algorithm based on neural collaborative filtering [4,5], that supports a systematic and fast evaluation of MV health status for medical practitioners. Our method substantially extends our work published in the conference paper [6] and improves on the Robust Non-negative Matrix Factorization method (RNMF), an unsupervised segmentation method proposed in [7] with a threefold contribution: (i) we use a neural collaborative filtering technique [5,4] that generalizes the matrix factorization and accounts for both linear and non-linear contributions of the myocardial wall motion, in combination with a parametrized threshold operator to learn the high dimensional sparse signal that captures the MV; (ii) we leverage the information of both the optical flow of the sparse signal and of the low dimensional time series representation of the echo to delineate the region of interest (ROI); (iii) we apply post-processing algorithms to improve the final MV segmentation. The method outperforms RNMF on a dataset of 39 patients affected with MV dysfunction and mitral regurgitation, and on an additional independent public dataset of 46 patients extracted from the EchoNet-Dynamic dataset [8].…”

“…The proposed segmentation model is composed of many stages and follows the structure of other unsupervised methods (see [6,7,11] and the literature review in Section 3). First, the echo video is embedded in a low dimensional space using a factorization technique (e.g.…”

Privacy-Preserving Artificial Intelligence Techniques in Biomedicine

“…There have also been a couple of attempts at incorporating federated learning into multitask learning and transfer learning in general. 138 139 140 However, to the best of our knowledge, FedHealth 135 is the only federated transfer learning framework specifically designed for health care applications. It enables users to train personalized models for their wearable health care devices by aggregating the data from different organizations without compromising privacy.…”

Mitral Valve Segmentation Using Robust Nonnegative Matrix Factorization

“…For an overview of the optimizations, we refer here to [34,35]. To improve the decomposition results of NMF, among others, orthogonality condition [36] or sparsity conditions [26,37,38] were added to NMF; in this paper, we consider the sparsity constraint of [26], known as RNMF, which is a special case of a method used by Corinzia et al [39], who used neural network matrix factorization combined with a threshold network, which is trained in an unsupervised manner, on each echocardiographic video individually to capture the motion of the mitral valve. In order to define the region of the heart valve, they additionally localized the heart valve region, using optical flow calculated on the outcome of the neural network.…”