acceptodds
Withdrawn from ICLR 2027

PathoField: Capacity-Controlled Feature-Space Residual Adaptation for Few-Shot Healthy-to-Pathological 3T-to-7T MRI Synthesis

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Paired pathological 3T–7T MRI is scarce, making adaptation of high-capacity synthesis models difficult in the few-shot regime. We introduce PathoField, a capacity-controlled transfer strategy that freezes a healthy-pretrained 3T-to-7T generator and learns four residual bottleneck adapters in decoder feature space. PathoField exposes only 45,696 trainable parameters (0.35% of the 13.20M-parameter backbone), approximately fewer than unrestricted full fine-tuning, while preserving the pretrained mapping exactly at initialization. We further characterize the asymmetric optimization induced by zero initialization and derive a local capacity-dependent generalization analysis. On paired Penn epilepsy data, all reported test checkpoints are re-evaluated with a common all-slice spatial-SSIM/PSNR protocol. At , PathoField reaches 0.7941 SSIM / 21.15,dB with 45,696 trainable parameters, compared with 0.8015 / 21.12,dB for Full-FT with 13.25M parameters and 0.7117 / 21.10,dB for rank-3 Conv-LoRA with 54,912 parameters. Thus, on this fixed split, PathoField comes within 0.0074 SSIM of unrestricted fine-tuning while improving SSIM by 0.0824 over the comparable-budget weight-space baseline. A separate 23-subject frozen-transfer audit yields 0.4731 SSIM / 15.11,dB. These results reveal a data-dependent fidelity–capacity trade-off: strong capacity restriction incurs a clear cost in the lowest-data regimes, but at , PathoField recovers most of Full-FT structural fidelity with approximately fewer trainable parameters.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
medical image synthesis, 3T-to-7T MRI, parameter-efficient fine-tuning, few-shot learning, domain adaptation, image-to-image translation, magnetic resonance imaging, feature-space adaptation, low-rank adaptation, generative models

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_dQLxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

PathoField adapts a healthy-pretrained 3T-to-7T MRI generator to a pathological cohort by freezing the backbone and training four zero-initialized residual bottleneck adapters. The paper investigates the resulting capacity–fidelity trade-off through comparisons with unrestricted fine-tuning and Conv-LoRA, together with a local capacity-dependent generalization analysis.

Critical strengths

At k=16, PathoField achieves competitive performance for Full-FT while training fewer parameters. It also exceeds the comparable-budget Conv-LoRA baseline, which provides an encouraging operating point for compact target-domain adaptation.

Critical weaknesses
  1. The methodological novelty and explanatory depth is limited. The method mainly relies on standard bottleneck adapters, with limited methodological novelty. The theoretical analysis also fails to clearly explain why feature-space adaptation performs better than weight-space adaptation.

  2. The writing and presentation need improvement, as several sections are difficult to follow and affect the overall clarity of the paper.

Questions for authors
  1. PathoField performs worse than Scratch and Full-FT at (k=1) and (k=8), while Scratch still outperforms it at (k=16). What is the actual advantage of the proposed method for few-shot adaptation?

  2. PathoField shows a sharp improvement from (k=8) to (k=16), while other methods remain relatively stable. This unusual behavior requires further explanation.

  3. Identity-preserving initialization is already common in adapter-based methods. What makes this design fundamentally different from existing approaches?

  4. The theoretical analysis mainly relates generalization to parameter count, without explaining the observed advantage over LoRA. What new insight does this analysis actually provide?

  5. The paper emphasizes parameter efficiency but provides no actual runtime or memory measurements. What practical efficiency gains does PathoField offer beyond reducing trainable parameters?

Reviewer_fQBAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PathoField adapts a healthy-pretrained 3T-to-7T MRI generator to epilepsy data using four decoder bottleneck adapters. It trains only 45.7k parameters and studies the trade-off between adaptation capacity and reconstruction fidelity.

Critical strengths

The reduction in trainable parameters is substantial. The main comparisons use a common evaluation protocol and include a similarly sized Conv-LoRA baseline. The paper also clearly acknowledges performance degradation in the smallest-data settings.

Critical weaknesses
  1. Limited methodological novelty. The method uses standard residual bottleneck adapters. The local capacity bound does not explain why feature-space adaptation should outperform weight-space adaptation (Sections 2–4).
  2. Narrow empirical support. Adaptation results rely on one split with only four test subjects. PathoField underperforms both Scratch and Full-FT at k=1k=1 and k=8k=8; its strongest result occurs at k=16k=16, where only one Conv-LoRA configuration is evaluated (Tables 3–4).
  3. Inconsistent evaluation of the central hypothesis. The rank–data ablation still uses the legacy similarity metric, whereas the main results use spatial SSIM. Consequently, the reported capacity trends remain unverified under the final evaluation protocol (Table 9).
  4. Pathological fidelity remains untested. Whole-image SSIM/PSNR cannot establish preservation of subtle abnormalities. Without lesion-level evaluation or controls for site differences, the study cannot determine whether adaptation addresses pathology or broader acquisition differences (Sections 5 and 7).
Questions for authors
  1. Does the advantage over Conv-LoRA persist across repeated subject splits and alternative configurations with comparable parameter budgets?
  2. Do the rank–data conclusions hold when the entire sweep is evaluated and checkpoints are selected using spatial SSIM?
  3. What evidence distinguishes pathology-related adaptation from correction of site or acquisition differences?
Reviewer_HqUyOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the challenging few‑shot 3T‑to‑7T MRI synthesis task under pathological conditions, where paired pathological 3T‑7T samples are extremely scarce. The authors propose PathoField, a feature‑space residual bottleneck adapter based parameter‑efficient fine‑tuning (PEFT) strategy. The approach freezes a well‑pretrained healthy‑domain 3T‑to‑7T U‑Net generator and only optimizes four zero‑initialized residual bottleneck adapters inserted in decoder feature branches. Only 45,696 parameters (0.35% of full backbone) are trainable, while the pretrained mapping is strictly preserved at initialization.

Critical strengths

Zero‑initialization of up‑projection weights guarantees exact identity mapping right after adapter insertion, so the source pretrained mapping is not corrupted at the beginning of adaptation. The paper theoretically analyzes the asymmetric first‑step update dynamics, which is a nice mechanistic insight distinguishing this adapter from many existing PEFT modules.

Critical weaknesses
  1. The paper only counts raw FP32 trainable parameter payload. It does not report GPU peak memory, inference runtime / throughput. For medical deployment scenarios, frozen backbone plus inserted adapters may introduce extra feature‑branch overhead. Readers cannot tell how PathoField compares against Full‑FT / Conv‑LoRA in actual runtime and memory consumption. Please add hardware‑matched profiling (e.g., on 256×256 slices, batch size 1) for representative methods.

  2. The current evaluation is only global slice‑wise SSIM / PSNR. Pathology‑relevant metrics (lesion region SSIM/PSNR) are missing due to annotation‑grid misalignment. Since the target task is epilepsy‑relevant pathological MRI synthesis, the model’s capability to preserve pathological lesion structure cannot be validated. The authors should discuss potential risks: high global SSIM does not guarantee clinically meaningful lesion preservation. If lesion masks can be aligned in revision, add lesion‑restricted metrics; otherwise add more qualitative lesion‑focused visual examples and elaborate clinical limitations in discussion / limitation sections.

  3. The derived generalization bound holds under local assumptions (fixed ReLU gates, Lipschitz downstream mapping, bounded parameter neighbourhood). It serves as a “capacity lens” rather than a predictive quantitative bound. Some text could accidentally mislead readers to treat it as a global guarantee for pathology domain shift. Please revise wording throughout Section 4 to repeatedly emphasize its local nature and underlying assumptions.

Questions for authors

The derived generalization bound holds under local assumptions (fixed ReLU gates, Lipschitz downstream mapping, bounded parameter neighbourhood). It serves as a “capacity lens” rather than a predictive quantitative bound. Some text could accidentally mislead readers to treat it as a global guarantee for pathology domain shift. Please revise wording throughout Section 4 to repeatedly emphasize its local nature and underlying assumptions.

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

Related papers

Discussion (0)

Sign in to comment.