Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LaTiM: Longitudinal representation learning in continuous-time models to predict disease progression

Rachid Zeghlache, Pierre-Henri Conze, Mostafa El Habib Daho, Yihao Li, Hugo Le Boité, Ramin Tadayoni, Pascal Massin, Béatrice Cochener, Alireza Rezaei, Ikram Brahim, Gwenolé Quellec, Mathieu Lamard

TL;DR

The paper addresses predicting disease progression from irregular longitudinal medical images by integrating neural ODEs with self-supervised pretraining. It introduces a time-aware head within a SIMCLR/BYOL-inspired SSL framework to condition predictions on time, using novel time augmentations and a disease progression-aligned objective. Empirical results on the OPHDIAT diabetic retinopathy dataset show significant improvements in AUC and Kappa over baselines, with BYOL-based pretraining and recurrent NODE architectures performing best, and the approach also stabilizing NODE training. This time-aware, SSL-guided pretraining framework offers a flexible, continuous-time representation for longitudinal medical imaging and could generalize to other diseases and temporal modeling tasks.

Abstract

This work proposes a novel framework for analyzing disease progression using time-aware neural ordinary differential equations (NODE). We introduce a "time-aware head" in a framework trained through self-supervised learning (SSL) to leverage temporal information in latent space for data augmentation. This approach effectively integrates NODEs with SSL, offering significant performance improvements compared to traditional methods that lack explicit temporal integration. We demonstrate the effectiveness of our strategy for diabetic retinopathy progression prediction using the OPHDIAT database. Compared to the baseline, all NODE architectures achieve statistically significant improvements in area under the ROC curve (AUC) and Kappa metrics, highlighting the efficacy of pre-training with SSL-inspired approaches. Additionally, our framework promotes stable training for NODEs, a commonly encountered challenge in time-aware modeling.

LaTiM: Longitudinal representation learning in continuous-time models to predict disease progression

TL;DR

The paper addresses predicting disease progression from irregular longitudinal medical images by integrating neural ODEs with self-supervised pretraining. It introduces a time-aware head within a SIMCLR/BYOL-inspired SSL framework to condition predictions on time, using novel time augmentations and a disease progression-aligned objective. Empirical results on the OPHDIAT diabetic retinopathy dataset show significant improvements in AUC and Kappa over baselines, with BYOL-based pretraining and recurrent NODE architectures performing best, and the approach also stabilizing NODE training. This time-aware, SSL-guided pretraining framework offers a flexible, continuous-time representation for longitudinal medical imaging and could generalize to other diseases and temporal modeling tasks.

Abstract

This work proposes a novel framework for analyzing disease progression using time-aware neural ordinary differential equations (NODE). We introduce a "time-aware head" in a framework trained through self-supervised learning (SSL) to leverage temporal information in latent space for data augmentation. This approach effectively integrates NODEs with SSL, offering significant performance improvements compared to traditional methods that lack explicit temporal integration. We demonstrate the effectiveness of our strategy for diabetic retinopathy progression prediction using the OPHDIAT database. Compared to the baseline, all NODE architectures achieve statistically significant improvements in area under the ROC curve (AUC) and Kappa metrics, highlighting the efficacy of pre-training with SSL-inspired approaches. Additionally, our framework promotes stable training for NODEs, a commonly encountered challenge in time-aware modeling.
Paper Structure (9 sections, 6 equations, 2 figures, 1 table)

This paper contains 9 sections, 6 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. SimCLR-based approach: disease progression alignment
  4. BYOL-based approach: temporal evolution and temporal consistency
  5. Experiments and results
  6. Dataset and implementation details.
  7. Results
  8. Ablation study
  9. Conclusion

Figures (2)

  • Figure 1: a), c) refer to SimCLR SimCLR and BYOL grill2020bootstrap, and e) refer to the standard paradigm for predicting disease progression at a certain point in the future, using a single image. b), d), and f) adapt the common SSL paradigm and supervised classification to disease progression with the introduction of a time-aware head.
  • Figure 2: Proposed augmentation techniques that we employed to mimic the popular SSL paradigms (SimCLR SimCLR on the left, BYOL grill2020bootstrap on the right) with neural ODE.