acceptodds
Withdrawn from ICLR 2027

Beyond One-Step Residuals: Trajectory-Aware Manifold Projection for Few-Shot Medical Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visual foundation models provide structured normal feature spaces for manifold-based anomaly detection. However, in few-shot medical settings, inter-individual variation induces train-test distribution shifts, leaving limited normal samples insufficient to characterize the normal manifold. Inspired by a dynamical systems perspective, we propose Trajectory-Aware Manifold Projection (TAMP), which exploits trajectory information beyond initial projection residuals. TAMP learns normal-manifold projectors through consistency-guided multi-level learning and repeatedly applies them during inference to construct recovery trajectories. We observe that initial residuals capture deviations, while subsequent projection steps reveal different responses to normal variations and lesions. Using consecutive projection displacements as additional anomaly cues, we introduce the post-projection trajectory ratio and terminal convergence residual to measure relative continued changes after the initial projection and the remaining final-step update, respectively. These complement the initial residual for anomaly scoring. Experiments on multiple medical benchmarks demonstrate improved detection and localization over existing few-shot methods, with competitive results against full-shot unsupervised methods using only a few normal references.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Medical Anomaly Detection, Few-Shot Learning, Manifold Projection, Projection Trajectories, Visual Foundation Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_unAMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TAMP, a Trajectory-Aware Manifold Projection method, to address the insufficient coverage of normal samples in few-shot medical anomaly detection. The method trains multi-layer feature projectors using normal images and synthetic anomalous images, and uses feature recovery trajectories formed through repeated projections as evidence of anomalies to assist anomaly detection. The final anomaly score combines the initial projection residual, trajectory change ratio, and terminal residual. The paper conducts experiments on four medical datasets.

Critical strengths
  1. The paper focuses on the problem of insufficient normal sample coverage in few-shot medical image anomaly detection and uses the continued displacement after the initial residual as supplementary evidence. The motivation is clear and directly corresponds to the single-step projection baseline.

  2. The paper conducts experiments on four medical benchmarks and includes five random runs, as well as ablation studies on the number of projection steps, trajectory components, training losses, and projector architectures, providing relatively comprehensive experimental results.

  3. The initial deviation, trajectory ratio, and terminal residual characterize the initial state, evolution, and convergence state, respectively. The scoring formulation is intuitive, with manageable computational overhead.

Critical weaknesses
  1. The projector is trained only with a single-step reconstruction objective, without imposing contraction, idempotence, or fixed-point constraints, nor are multi-step trajectories unrolled during training. Therefore, the persistent displacement observed after repeated applications may arise from the projector's own instability or out-of-distribution iteration errors. Although the paper cites stable manifold theory, it does not verify the Jacobian spectral radius, the convergence region, or the stable fixed-point conditions for normal samples. Thus, the explanation of the core mechanism remains insufficiently supported.

  2. The comparison with full-shot medical anomaly detection methods mixes different backbones, training data volumes, and implementation protocols, making it insufficient to support the strong claim that “four-shot outperforms full-shot.” Moreover, pixel-level evaluation is conducted only on RESC.

  3. The trajectory ratio uses the initial residual as its denominator, which may lead to amplified values when the initial residual of normal samples is very small. The paper does not provide a stabilization term, a clipping strategy, or a corresponding sensitivity analysis.

Questions for authors

Please report multi-step error curves for normal and anomalous samples, the spectral radius of the projector's Jacobian or Lipschitz estimates, and compare single-step training, unrolled multi-step training, explicit idempotence regularization, and contraction regularization to demonstrate that the trajectory signal originates from lesions rather than iterative instability.

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

This paper proposes Trajectory-Aware Manifold Projection (TAMP) for few-shot medical anomaly detection. The authors observe that conventional one-step projection residuals may struggle to distinguish normal variations from pathological anomalies due to limited normal-manifold coverage. To address this issue, TAMP learns multi-level normal-manifold projectors and exploits feature evolution across multiple projection steps. Specifically, the method introduces a post-projection trajectory ratio and a terminal convergence residual to capture complementary anomaly evidence beyond the initial projection residual. Experiments on multiple medical benchmarks demonstrate promising performance in anomaly detection and localization under few-shot settings.

Critical strengths

This paper explores a novel perspective for distinguishing normal and anomalous regions in medical images by analyzing the evolution of feature residuals over multiple manifold projection steps. The authors observe that projection responses in normal regions tend to diminish rapidly, whereas anomalous regions often exhibit more persistent feature updates. Based on this observation, the authors propose Trajectory-Aware Manifold Projection (TAMP) to exploit multi-step projection trajectories as additional anomaly evidence. The overall idea is interesting, and the experimental results are promising.

Critical weaknesses

1.Distribution gap between synthetic and real anomalies. The authors employ CutPaste to generate synthetic anomalies during training and learn normal feature recovery mappings. However, such synthetic perturbations may differ substantially from real medical anomalies in terms of morphology, texture, and semantic characteristics. It remains unclear whether this distribution gap affects the learned projectors' behavior when processing real pathological features.

2.Insufficient justification and analysis of the projection steps. The authors set the number of projection steps to (T=3). Although Figure 6 compares detection performance and inference time under different projection steps, the manuscript provides limited analysis of how the residual evolution and discriminative behavior of normal and anomalous regions change as the number of projection steps increases.

3.Missing ground-truth annotations in qualitative visualizations. Figure 5 presents anomaly heatmaps generated by different methods but does not include the corresponding ground-truth annotations. This makes it difficult to visually assess and compare their anomaly localization accuracy.

4.Insufficient comparisons with recent medical anomaly detection methods. Although the authors compare their method with general-purpose anomaly detection approaches, such as WinCLIP, and include several medical anomaly detection baselines, comparisons with some recent medical anomaly detection methods, such as MVFA-AD and MediCLIP, are still missing. 5.Code availability and reproducibility. The current submission does not provide accessible source code.

Questions for authors

See above.

Reviewer_MLxaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TAMP detects and localizes anomalies from one, two or four normal reference images. It trains projectors on multiple levels of frozen DINOv3 features using synthetic perturbations, reconstruction loss, adaptive fusion, and cross level consistency. Training uses single step reconstruction and inference repeatedly applies the learned projectors. The paper combines initial residuals, normalized sum of displacements from later projection steps and final step residual to produce image scores and patch maps. The paper evaluates detection on HeadCT, BrainMRI, and ZhangLab CXR and anomaly localization on RESC. It reports improvements over few shot baseline and ablations are presented to show the contribution of multi-level learning and trajectory based scoring.

Critical strengths
  1. Four shot TAMP reports 94.8 average image level AUROC versus 89.9 for the strongest competing method. On one shot RESC it reports 88.4 PRO versus 81.0. Both seem to be meaningful improvements across anomaly detection and localization.

  2. Table 5 reports about 90.2 average AUROC without trajectory scoring, 92.1 with endpoint displacement, and 94.8 with the proposed score. It also includes ablations of individual trajectory components. These suggest that repeatedly applying projectors trained for single step reconstruction provides useful additional anomaly information. TAMP extends FoundAD's projector approach through multi level learning and consecutive projection update.

Critical weaknesses
  1. In the appendix B, paper says that T = 3 was selected through ablation, while beta, gamma, and top kappa pooling were chosen separately for each dataset. However, the paper does not specify the validation data, selection criteria, or labels used. This matters because Table 8 shows that BrainMRI AUROC drops from 96.1 to 86.9 under different parameter settings. Without a clear selection protocol independent of the test results, it is difficult to assess how reliably the reported improvements would generalize.

  2. Appendix A defines fusion inputs using reconstruction errors against paired clean targets, but Eq. 8 does not explain how these weights are computed at inference when clean targets are unavailable. This leaves an important part of the inference procedure unclear.

  3. In the table 5 it does not compare the proposed trajectory score against a simpler combination of the same step wise displacements. As a result, it is unclear whether the specific scoring formulation provides a meaningful advantage over simpler alternatives.

Questions for authors
  1. Can you specify the objectives, data, labels, and final settings used to select beta, gamma, kappa, and T. Were the baselines given the same validation information ?

  2. How are the fusion inputs computed and normalized at inference when paired clean targets are unavailable? Could the authors report the learned weight distributions during training and for normal and abnormal test images? How does adaptive fusion compare with fixed weights when using the same trained projectors and scoring procedure?

  3. How does the proposed score compare with a simpler combination of the same projection updates, such as d0+α∑t=1T−1dtd^0 + \alpha \sum_{t=1}^{T-1} d^t under identical experimental settings? Also it would be appreciated to explain why Baseline + MF and TAMP without AlA_l report identical results in Table 5, and whether they use the same training objectives and projector checkpoints?

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

Related papers

Discussion (0)

Sign in to comment.