Source-linked AI summary
Deep learning cardiac motion analysis for human survival prediction
Ghalib A. Bello, Timothy J. W. Dawes, Jinming Duan, Carlo Biffi, Antonio de Marvao, Luke S. G. E. Howard, J. Simon R. Gibbs, Martin R. Wilkins, Stuart A. Cook, Daniel Rueckert, Declan P. O'Regan
TL;DR
The paper addresses survival prediction from high-dimensional, time-resolved cardiac motion data. It uses automated 3D motion modelling and 4Dsurvival to learn survival-relevant representations, achieving higher Harrell's C-index than conventional human-derived indices.
Problem
High-dimensional cardiac motion trajectories require accurate tracking and efficient representations for predicting time-to-event outcomes in patients with pulmonary hypertension.
Method
A fully convolutional network creates time-resolved 3D cardiac segmentations, which feed a supervised denoising autoencoder called 4Dsurvival trained with a Cox survival objective.
Results
Harrell's C-index was 0.73 (95% CI: 0.68-0.78) for 4Dsurvival versus 0.59 (95% CI: 0.53-0.65) for the conventional parameter model, p < .0001.
Takeaways & Limitations
Compressed representations of dynamic 3D heart motion can support automated and interpretable survival prediction from clinical images.
Takeaways & Limitations
The Cox model assumes that each subject's hazard ratio relative to baseline remains constant over time.
Abstract
from arXiv · showhide
Motion analysis is used in computer vision to understand the behaviour of moving objects in sequences of images. Optimising the interpretation of dynamic biological systems requires accurate and precise motion tracking as well as efficient representations of high-dimensional motion trajectories so that these can be used for prediction tasks. Here we use image sequences of the heart, acquired using cardiac magnetic resonance imaging, to create time-resolved three-dimensional segmentations using a fully convolutional network trained on anatomical shape priors. This dense motion model formed the input to a supervised denoising autoencoder (4Dsurvival), which is a hybrid network consisting of an autoencoder that learns a task-specific latent code representation trained on observed outcome data, yielding a latent representation optimised for survival prediction. To handle right-censored survival outcomes, our network used a Cox partial likelihood loss function. In a study of 302 patients the predictive accuracy (quantified by Harrell's C-index) was significantly higher (p < .0001) for our model C=0.73 (95$\%$ CI: 0.68 - 0.78) than the human benchmark of C=0.59 (95$\%$ CI: 0.53 - 0.65). This work demonstrates how a complex computer vision task using high-dimensional medical image data can efficiently predict human survival.
Introduction
The paper addresses survival prediction from complex 3D cardiac motion in pulmonary hypertension, where individual risk stratification matters. It combines automated cardiac segmentation with learned motion representations to compare against conventional human-derived volumetric indices.
- Pulmonary hypertension involves right ventricular dysfunction, high mortality, and treatment decisions that depend on individual risk stratification.
- Cardiac motion is a complex, non-linear, rhythmic process that reflects disease-related changes in geometry and function.
- The study tests whether learned features of complex 3D cardiac motion can improve prognostic accuracy over explicit myocardial motion measurements.
- The method uses a fully convolutional network trained from manual anatomical priors to generate smooth 3D frame-wise cardiac motion renderings.
- These motion renderings feed 4Dsurvival, a supervised denoising autoencoder designed to learn noise-robust latent representations salient for survival prediction.
- The learned model is compared with conventional survival prediction based on human-derived volumetric indices.
Results
The study analyzed 302 patients with incident pulmonary hypertension using automated, time-resolved cardiac motion models. 4Dsurvival produced motion-based risk predictions, distinguished survival groups, and yielded interpretable latent and right-ventricular regional visualizations.
- Study population: 302 patients with incident pulmonary hypertension were included, with 28% all-cause mortality during follow-up.Patients were investigated between 2004 and 2017 and followed until November 27, 2017.
- Motion processing: 69,820 label maps were generated by automatically segmenting the ventricles across slices and 20 temporal phases.Image registration tracked corresponding anatomical points before constructing dense cardiac motion models.
- Predictive performance: C = 0.73 (95% CI: 0.68-0.78) for 4Dsurvival versus C = 0.59 (95% CI: 0.53-0.65) for the conventional parameter model after bootstrap optimism correction.The difference was statistically significant at p < .0001.
- Risk stratification: χ2 = 20.7, p < .0001 for 4Dsurvival versus χ2 = 5.7, p = .0173 for the conventional parameter model in log-rank comparisons of median-defined risk groups.Kaplan-Meier estimates were shown with 95% confidence intervals.
- Interpretability: Right-ventricular saliency maps identified contributions from spatially distant but functionally synergistic regions to survival prediction in pulmonary hypertension.Regional contributions were estimated from vertex-level associations between motion magnitude and predicted risk score.
Discussion
The paper argues that compressed representations of dynamic cardiac motion can support time-to-event analysis beyond conventional clinical variables and global contraction measures. It combines time-resolved cardiac segmentation with supervised representation learning and identifies larger, multicentre validation as a next step.
- Compressed representations of dynamic biological systems moving in 3D space offer an approach to time-to-event analysis.
- Conventional linear regression and global contraction measures do not capture the dynamic physiological complexity of heart disease.
- Cardiac time-domain segmentation with fully convolutional networks has achieved expert-level performance, while motion-model components show prognostic utility.
- The method extends denoising autoencoders with a prediction branch so learned representations support survival-related discrimination as well as reconstruction.
- Larger and more diverse multicentre patient groups, followed by external clinical validation against established risk scores, are identified as implementation steps.
Methods
The study combines automated biventricular segmentation and motion registration with a denoising autoencoder and Cox model to represent cardiac motion for survival prediction.
- Segmentation: A multi-task fully convolutional network learned ventricular segmentation and landmark locations from manually labelled CMR volumes.The objective combined region-associated, landmark-associated, and weight-decay terms, with stochastic-gradient optimization.
- Motion modeling: Atlas-based non-rigid registration converted the segmentations into anatomically smooth, high-resolution motion models with corresponding vertices across temporal frames and subjects.Each subject was represented by 20 surface meshes, each containing 18,028 anatomically corresponding vertices.
- Prediction network: The 4Dsurvival network used a denoising autoencoder to compress cardiac motion into a low-dimensional latent code linked to a survival predictor.Random masking corrupts the input during encoding, while the latent representation is trained to reconstruct motion and support survival prediction.
- Survival modeling and validation: The survival component used a Cox proportional hazards model, whose hazard-ratio assumption remains constant with respect to time.Hyperparameters were selected using particle swarm optimization, and model performance was assessed with Harrell’s concordance index and bootstrap optimism correction.
Author contributions statement (CRediT)
The paper assigns contributions across methodology, software, analysis, writing, investigation, supervision, conceptualization, funding acquisition, and final manuscript review.
- Contributions span methodology, software, formal analysis, writing, investigation, supervision, conceptualization, funding acquisition, and manuscript review.