CRONOS: Continuous Time Reconstruction for 4D Medical Longitudinal Series
Nico Albert Disch, Saikat Roy, Constantin Ulrich, Yannick Kirchhoff, Maximilian Rokuss, Robin Peretzke, David Zimmerer, Klaus Maier-Hein
TL;DR
CRONOS introduces a novel continuous-time framework for 4D medical imaging prediction, enabling voxel-level forecasting from multiple past scans at arbitrary times, improving accuracy and flexibility over existing methods.
Contribution
It is the first model to perform continuous sequence-to-image forecasting for 3D medical data, supporting both discrete and continuous timestamps in a unified approach.
Findings
Outperforms baseline models on three public datasets.
Supports arbitrary time predictions in 3D medical imaging.
Operates efficiently in voxel space.
Abstract
Forecasting how 3D medical scans evolve over time is important for disease progression, treatment planning, and developmental assessment. Yet existing models either rely on a single prior scan, fixed grid times, or target global labels, which limits voxel-level forecasting under irregular sampling. We present CRONOS, a unified framework for many-to-one prediction from multiple past scans that supports both discrete (grid-based) and continuous (real-valued) timestamps in one model, to the best of our knowledge the first to achieve continuous sequence-to-image forecasting for 3D medical data. CRONOS learns a spatio-temporal velocity field that transports context volumes toward a target volume at an arbitrary time, while operating directly in 3D voxel space. Across three public datasets spanning Cine-MRI, perfusion CT, and longitudinal MRI, CRONOS outperforms other baselines, while…
Peer Reviews
Decision·ICLR 2026 Poster
1. Longitudinal 4D imaging is central to disease progression analysis, yet remains underexplored by the ML community. 2. Extending FM from noise-to-sample to multi-context-to-target volumetric flows is conceptually creative and unifies discrete and continuous-time modeling within one ODE framework. 3. The discrete/continuous variants and Fourier-time conditioning offer a simple yet flexible approach adaptable to both regularly and irregularly sampled medical series.
1. The “velocity loss” $\|v_\theta(X_\tau,T_\tau)-(I_{\text{target}}-I)\|^2$ treats voxel-intensity differences as flow supervision. This deviates from FM’s probabilistic transport interpretation and lacks theoretical grounding in physical or latent-space dynamics. 2. ConvLSTM, SimVP, and ViViT are optimized for long, dense 2D videos, not short 3D series. Their poor results may reflect mismatch rather than true inferiority. 3. The “continuous” setting is created by subsampling ACDC; no genuinely
1. Problem Relevance: The paper addresses a critical problem in 3D medical imaging by proposing a unified framework for continuous sequence-to-image forecasting. 2. Technical contribution: the method makes use of Flow Matching applied to the medical imaging domain through temporal broadcasting. The theoritical formulation is clear and well defined. 3. Performance. CRONOS shows strong performance over LCI baseline and the spatio-temporal baselines across all three datasets. 3. Evaluation: The
1. Lack of expert validation: only voxel-wise metrics (NRMSE, PSNR, SSIM) are provided to assess performance but no clinical expert validation is shown nor other relevant anatomical metrics. 2. Limited improvement: CRONOS scores 94.51% vs. LCI's 92.79% SSIM on ACDC, which appears to be a small gain given the added computational cost. 3. Methodological issues: there are concerns the LOCF (Last Observed Carry Forward) filling strategy (eq 19) may introduce artifacts. Even though it is demonstrat
Addresses an important, under-explored problem: The paper targets 3D medical image sequence forecasting from multiple prior scans at irregular time intervals – a challenging and clinically relevant task (e.g. disease progression) that has received limited attention. Unified continuous-time approach: CRONOS is a novel framework combining multiple-context inputs with continuous-time prediction in one model. It appears to be the first method to handle real-valued time stamps and multiple input vol
Limited algorithmic novelty (major): The approach offers minimal new methodology, essentially applying the existing Flow Matching paradigm to the many-to-one 3D forecasting setting without introducing new algorithmic contributions. The core idea (learning a velocity field via flow matching) is borrowed from prior work, with novelty lying primarily in its application to this domain rather than in the technique itself. Missing baseline comparisons (major): The evaluation lacks comparisons with se
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsMachine Learning in Healthcare · Generative Adversarial Networks and Image Synthesis · Radiomics and Machine Learning in Medical Imaging