Classification of red cell dynamics with convolutional and recurrent neural networks: a sickle cell disease case study.

The fraction of red blood cells adopting a specific motion under low shear flow is a promising inexpensive marker for monitoring the clinical status of patients with sickle cell disease. Its high-throughput measurement relies on the video analysis of thousands of cell motions for each blood sample to eliminate a large majority of unreliable samples (out of focus or overlapping cells) and discriminate between tank-treading and flipping motion, characterizing highly and poorly deformable cells respectively. Moreover, these videos are of different durations (from 6 to more than 100 frames). We present a two-stage end-to-end machine learning pipeline able to automatically classify cell motions in videos with a high class imbalance. By extending, comparing, and combining two state-of-the-art methods, a convolutional neural network (CNN) model and a recurrent CNN, we are able to automatically discard 97% of the unreliable cell sequences (first stage) and classify highly and poorly deformable red cell sequences with 97% accuracy and an F1-score of 0.94 (second stage). Dataset and codes are publicly released for the community.

Tropical Cyclone Track Forecasting Using Fused Deep Learning From Aligned Reanalysis Data.

The forecast of tropical cyclone trajectories is crucial for the protection of people and property. Although forecast dynamical models can provide high-precision short-term forecasts, they are computationally demanding, and current statistical forecasting models have much room for improvement given that the database of past hurricanes is constantly growing. Machine learning methods, that can capture non-linearities and complex relations, have only been scarcely tested for this application. We propose a neural network model fusing past trajectory data and reanalysis atmospheric images (wind and pressure 3D fields). We use a moving frame of reference that follows the storm center for the 24 h tracking forecast. The network is trained to estimate the longitude and latitude displacement of tropical cyclones and depressions from a large database from both hemispheres (more than 3,000 storms since 1979, sampled at a 6 h frequency). The advantage of the fused network is demonstrated and a comparison with current forecast models shows that deep learning methods could provide a valuable and complementary prediction. Moreover, our method can give a forecast for a new storm in a few seconds, which is an important asset for real-time forecasts compared to traditional forecasts.

A rule-based method to model myocardial fiber orientation in cardiac biventricular geometries with outflow tracts.

Rule-based methods are often used for assigning fiber orientation to cardiac anatomical models. However, existing methods have been developed using data mostly from the left ventricle. As a consequence, fiber information obtained from rule-based methods often does not match histological data in other areas of the heart such as the right ventricle, having a negative impact in cardiac simulations beyond the left ventricle. In this work, we present a rule-based method where fiber orientation is separately modeled in each ventricle following observations from histology. This allows to create detailed fiber orientation in specific regions such as the endocardium of the right ventricle, the interventricular septum, and the outflow tracts. We also carried out electrophysiological simulations involving these structures and with different fiber configurations. In particular, we built a modeling pipeline for creating patient-specific volumetric meshes of biventricular geometries, including the outflow tracts, and subsequently simulate the electrical wavefront propagation in outflow tract ventricular arrhythmias with different origins for the ectopic focus. The resulting simulations with the proposed rule-based method showed a very good agreement with clinical parameters such as the 10 ms isochrone ratio in a cohort of nine patients suffering from this type of arrhythmia. The developed modeling pipeline confirms its potential for an in silico identification of the site of origin in outflow tract ventricular arrhythmias before clinical intervention.

Transfer Learning From Simulations on a Reference Anatomy for ECGI in Personalized Cardiac Resynchronization Therapy.

GOAL: Noninvasive cardiac electrophysiology (EP) model personalisation has raised interest for instance in the scope of predicting EP cardiac resynchronization therapy (CRT) response. However, the restricted clinical applicability of current methods is due in particular to the limitation to simple situations and the important computational cost. METHODS: We propose in this manuscript an approach to tackle these two issues. First, we analyze more complex propagation patterns (multiple onsets and scar tissue) using relevance vector regression and shape dimensionality reduction on a large simulated database. Second, this learning is performed offline on a reference anatomy and transferred onto patient-specific anatomies in order to achieve fast personalized predictions online. RESULTS: We evaluated our method on a dataset composed of 20 dyssynchrony patients with a total of 120 different cardiac cycles. The comparison with a commercially available electrocardiographic imaging (ECGI) method shows a good identification of the cardiac activation pattern. From the cardiac parameters estimated in sinus rhythm, we predicted five different paced patterns for each patient. The comparison with the body surface potential mappings (BSPM) measured during pacing and the ECGI method indicates a good predictive power. CONCLUSION: We showed that learning offline from a large simulated database on a reference anatomy was able to capture the main cardiac EP characteristics from noninvasive measurements for fast patient-specific predictions. SIGNIFICANCE: The fast CRT pacing predictions are a step forward to a noninvasive CRT patient selection and therapy optimisation, to help clinicians in these difficult tasks.

Noninvasive Personalization of a Cardiac Electrophysiology Model From Body Surface Potential Mapping.

GOAL: We use noninvasive data (body surface potential mapping, BSPM) to personalize the main parameters of a cardiac electrophysiological (EP) model for predicting the response to different pacing conditions. METHODS: First, an efficient forward model is proposed, coupling the Mitchell-Schaeffer transmembrane potential model with a current dipole formulation. Then, we estimate the main parameters of the cardiac model: activation onset location and tissue conductivity. A large patient-specific database of simulated BSPM is generated, from which specific features are extracted to train a machine learning algorithm. The activation onset location is computed from a Kernel Ridge Regression and a second regression calibrates the global ventricular conductivity. RESULTS: The evaluation of the results is done both on a benchmark dataset of a patient with premature ventricular contraction (PVC) and on five nonischaemic implanted cardiac resynchonization therapy (CRT) patients with a total of 21 different pacing conditions. Good personalization results were found in terms of the activation onset location for the PVC (mean distance error, MDE = 20.3 mm), for the pacing sites (MDE = 21.7 mm) and for the CRT patients (MDE = 24.6 mm). We tested the predictive power of the personalized model for biventricular pacing and showed that we could predict the new electrical activity patterns with a good accuracy in terms of BSPM signals. CONCLUSION: We have personalized the cardiac EP model and predicted new patient-specific pacing conditions. SIGNIFICANCE: This is an encouraging first step towards a noninvasive preoperative prediction of the response to different pacing conditions to assist clinicians for CRT patient selection and therapy planning.