acceptodds
Withdrawn from ICLR 2027

KneeForce: Longitudinal Knee Osteoarthritis X-ray Prediction via Diffusion Forcing with Temporal Relational Alignment and Trajectory Selection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Knee Osteoarthritis (KOA) is a serious chronic disease leading to disability, and anticipating its progression is essential for timely, personalized intervention. Existing prediction methods largely forecast non-image indicators such as future Kellgren-Lawrence (KL) grades or pain scores. However, these scalar outcomes provide only coarse, subjective assessments and lack the fine-grained visual details that clinicians need to guide treatment. Directly predicting future radiographs offers a more objective and interpretable alternative. Yet longitudinal forecasting remains challenging. Existing generative approaches struggle to maintain anatomical consistency over long horizons due to error accumulation and rigid inference. Moreover, they overlook the irregular sampling intervals inherent to clinical follow-up. To overcome these challenges, we propose KneeForce, a novel framework that adapts Diffusion Forcing for continuous disease trajectory modeling. Our method explicitly encodes the time interval to accommodate unevenly sampled inputs, and incorporates Temporal Relational Alignment, which leverages a DINO-based knee foundation model as a semantic expert to guide the generative process toward high-level disease patterns rather than mere texture synthesis. Additionally, a Trajectory Selection strategy is carefully designed to filter stochastic outputs, identifying the most clinically probable progression path. Extensive validation on the public OAI dataset and an external in-house dataset demonstrates that KneeForce consistently outperforms state-of-the-art medical and video synthesis methods in both radiographic fidelity and clinical grading accuracy. A blinded study by experienced clinicians further confirms the clinical reliability of our predictions, offering trustworthy visual evidence to support long-term prognosis.

Primary area:
generative models
Keywords:
Longitudinal Medical Image Prediction, Diffusion Forcing, Knee Osteoarthritis, Disease Progression Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_gnqnOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript proposes a diffusion forcing-based framework designed for longitudinal Knee OA radiographic prediction. The design is with three major components to adapt the diffusion forcing to this specific application. First, the time-index is encoded into the latent state to accommodate irregular clinical follow-up schedules. Second, temporal relation alignment distills structural prior knowledge from a frozen pre-trained expert model by aligning frame-to-frame similarity matrices. Third, based on the random nature of the diffusion setting, trajectory candidates are created and selected based on the specific criteria. The authors compare with the existing generation approaches and achieve the best results on two datasets (one public and one in-house) in image generation metrics and OA longitudinal grading accuracy.

Critical strengths
  • A reasonable engineering practice to adapt GenAI with longitudinal medical image prediction.
  • The problem is worthwhile for exploration and clinically meaningful.
Critical weaknesses
  • The overall framework looks incremental and lacks sufficient new insights to the domain. There are exisiting studies conducting knee OA image prediction in a longitudinal manner using diffusion models. Such generation results are further used for KL grading prediction.

  • The design of the proposed components is not well justified and verified.

Questions for authors
  • The choice of DF is considered by its general flexibility. I was wondering if there is any theoretical or empirical proof that it is superior to other relevant sequential generation strategies.

  • The delay dd controls the denoising process. Yet, its alignment with actual disease progression is not demonstrated. Using this to predict clinical-meaningful sequence is not well justified.

  • What evidence supports the candidate ranking and its selected output as the "most clinically probable progression"?

  • Missing "Direct-Forecasting" baseline: The paper lacks a comparison against classic direct prediction (e.g., using historical frames to directly predict future KL grades). I hope to see that the generated-image-based prediction can outperform this baseline to demonstrate the clinical values.

  • While an SSIM of 0.81 suggests the images look similar, in a clinical context, it still indicates noticeable differences (roughly 20% structural mismatch). The authors need to further analyze such differences and their clinical impacts, especially with the discussions of failure cases.

  • Can the authors provide an external-cohort ablation of time encoding and controlled experiments at follow-up intervals unseen during training?

  • How does performance vary by baseline severity, progression rate, and prediction horizon?

  • The human study only proves relative preference among different methods, not absolute clinical usability. It lacks the perspecitve whether generated images improve actual clinical decision-making.

  • The methods is with multiple components and with somehow complex designs. The reviewer has the concern regarding the reproducibility, while only the appendix states that 'For detailed implementation please refer to the source code.'

  • While SOTA, the absolute KL accuracy drops to ~62% at 96 months (nearly 40% of the predictions are misgraded and two-year observation has been conducted already). The image metric gains are marginal. These raise concerns about whether the generated images are reliable enough for real-world clinical use.

Reviewer_5CFTOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

KneeForce adapts Diffusion Forcing to longitudinal knee osteoarthritis X-ray prediction by integrating explicit time-interval encoding, a DINO-based Temporal Relational Alignment strategy, and a Trajectory Selection mechanism to handle unevenly sampled clinical follow-ups and generate clinically plausible disease progressions.

Critical strengths

The incorporation of explicit time-interval encoding addresses the non-uniform clinical follow-up schedules inherent in real-world longitudinal data, preventing miscalibrated progression rates.

Utilizing a DINO-based knee foundation model (OrthoFoundation) for Temporal Relational Alignment ensures that the diffusion model focuses on high-level pathological evolution (such as joint space narrowing) rather than superficial texture synthesis.

The Trajectory Selection strategy successfully mitigates generative stochasticity by filtering candidate trajectories to identify the most clinically probable progression path, which is further validated by a rigorous blinded reader study.

Critical weaknesses

The performance of Temporal Relational Alignment heavily relies on the representation capacity of the pre-trained DINO knee foundation model; any domain biases or errors present in the expert model risk propagating directly into the generative trajectories.

The requirement to generate and evaluate multiple candidate trajectories under varying schedules during inference introduces significant computational overhead, which could limit its practicality in real-time clinical workflows.

While validated on OAI and an in-house dataset, the framework's reliance on standard Kellgren-Lawrence grading and common progression patterns may constrain its ability to accurately predict atypical, rapid, or multi-factorial joint degeneration cases.

Questions for authors

See Weakness

Reviewer_eNidOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents KneeForce, a longitudinal knee osteoarthritis radiograph forecasting framework built on Diffusion Forcing. The method is evaluated on the OAI dataset and an external in-house cohort, with additional ablation studies and a blinded clinician evaluation. The problem is clinically meaningful, and the empirical evaluation is relatively comprehensive.

Critical strengths
  1. The paper addresses a clinically meaningful and relatively underexplored problem: predicting longitudinal KOA radiographic trajectories rather than only scalar KL grades or isolated future images.

  2. The use of Diffusion Forcing is well motivated for long-horizon sequence generation, and explicit time conditioning is a sensible design for irregular clinical follow-up. The TRA formulation using inter-frame relationships from a knee-specific foundation model is also technically coherent.

  3. The experimental evaluation is relatively strong for a medical generative modeling paper, including OAI, an external cohort, multiple longitudinal baselines, ablations, and a blinded clinician study.

Critical weaknesses
  1. KneeForce mainly combines Diffusion Forcing with time conditioning, representation alignment to a pretrained knee foundation model, and a separately trained trajectory-ranking module. Each component is reasonable, but the paper does not yet establish a sufficiently distinct learning principle beyond adapting existing techniques to longitudinal knee imaging. The distinction from representation-guided diffusion and prior irregular-time longitudinal generation methods should be made much sharper.

  2. Most OAI experiments follow a predefined acquisition schedule, and the paper itself shows that time encoding provides only marginal improvement in this setting.

  3. The external cohort is small and the reported results are averaged across available follow-ups rather than explicitly evaluating unseen target intervals. A stronger test would evaluate interpolation and extrapolation to acquisition times not observed during training.

  4. The ranking model is trained using the distance between generated trajectories and the ground-truth future sequence, and then selects among multiple diffusion schedules at inference. This is a reasonable selection strategy, but it introduces additional supervision and inference-time search. The paper should report random selection, the best fixed schedule, a simple image-quality-based selector, and an oracle upper bound to clarify how much of the final gain comes from the generator versus candidate selection.

Questions for authors

The author can consider responding to the comments in Weakness, and I will consider changing my rating.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.