acceptodds
Under review as a conference paper at ICLR 2027

Disentangled Learning Improves Implicit Neural Representations for Medical Reconstruction

Abstract

Implicit neural representations (INRs) have emerged as a powerful paradigm for medical imaging via physics-informed unsupervised learning. Classical INRs optimize an entire network from scratch for each subject, leading to inefficient training and suboptimal imaging quality. Recent initialization-based approaches attempt to inject population priors into pre-trained networks, yet they rely on high-quality images and update the prior-bearing parameters during fine-tuning. We present DisINR, a novel INR framework that explicitly disentangles shared and subject-specific representations. DisINR introduces a shared encoder–decoder pair and subject-specific encoders, whose features are jointly decoded for image reconstruction. By integrating differentiable forward models, it pre-trains the shared modules directly from limited raw measurements, removing the need for pre-acquired high-quality images. During test-time adaptation, only the subject-specific encoder is optimized, while the shared pair remains frozen, effectively preserving learned priors. Extensive evaluations on three representative medical imaging tasks show that DisINR significantly outperforms state-of-the-art INRs in both reconstruction accuracy and efficiency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Reconstruction, Unsupervised Learning, Implicit Neural Representations

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_p51MOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DisINR introduces a shared encoder–decoder pair together with subject-specific encoders, whose features are concatenated and jointly decoded to reconstruct the image. The shared pair is pre-trained across diverse subjects; during test-time adaptation it is frozen, and only a subject-specific encoder is optimized. The authors claim this effectively incorporates learned population priors into INR optimization, enabling faster and higher-quality reconstruction.

Critical strengths

Unlike STRAINER, DisINR places the shared pair (f_ϕ*, g_ψ*) where it can be frozen at test time, so that population priors are injected into the INR optimization while only the subject-specific encoder f_φ is optimized. This is a meaningful structural difference. The design yields the best reconstruction performance in PSNR and SSIM among the evaluated methods, consistently across the three tasks.

Critical weaknesses
  1. The paper credits its gain to freezing, on the grounds that freezing keeps the population prior. Two simple reference points would test this.

First, what happens with no prior at all? Keep the same architecture, remove the the population information, for example, start the prior encoder from random weights and still freeze it (not sure whether it is reasonable).

Second, what happens with only a trivial prior? Replace the shared feature with the average image of the training subjects. The paper states that the shared representation is “not simply the mean image,” but did not compare against the mean image.

These controls are inexpensive here. Because the input is a fixed coordinate grid and the shared modules are frozen, the shared branch produces the same output for every test subject, and the two branches are combined only by concatenation. Keeping one branch at a time would likewise show whether the shared branch carries real population information or merely keeps the frozen decoder within a working input range. The paper describes the shared representation as capturing “higher-level semantic structure,” but the term is never defined and no measurement supports it, and the decoder is not analyzed at all. Without these reference points there is no basis for judging how much the prior is actually worth.

  1. Is The shared pair pre-trained on a homogeneous cohort? Must the pre-training cohort match the test case in modality, anatomy, and contrast for the prior to be useful? If the shared pair were pre-trained over a diverse mix of subjects, would it still provide a useful prior? This matters for the claimed applicability across practical scenarios: if one shared pair is needed per anatomy and protocol, the pre-training cost must be paid repeatedly for different cohorts.
Questions for authors

Fundamentally, why should freezing the shared pair and fitting a subject-specific encoder be better than STRAINER’s shared-encoder/subject-decoder arrangement? Please make this argument explicit rather than resting it on the aggregate results.

It is seen from the results that fine-tuning the shared pair lowers the performance, probably due to the losing the prior, especially as finetune reuses the pre-training learning rate on already-converged weights with no stopping criterion. Can the author try intermediate settings: a small learning rate on the shared weights, a penalty for departing from the original shared pair of (f_ϕ*, g_ψ*) , or freezing the shared encoder while lightly adapting the decoder with the subject-specific encoder?

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

This paper proposes DisINR, an implicit neural representation framework for medical image reconstruction. It jointly learns shared and subject-specific representations through a measurement-domain loss, and freezes the shared encoder and decoder at test time while optimizing only the subject-specific encoder. Experiments on 3D volume fitting, undersampled MRI, and sparse-view CT demonstrate improved reconstruction quality and test-time efficiency over the evaluated INR baselines.

Critical strengths
  1. DisINR enables subject-specific adaptation while preserving shared representations by freezing the shared modules and optimizing only the subject-specific encoder. The freezing ablation further supports the reconstruction benefits of this design, demonstrating the practical value of separating shared and subject-specific learning.
  2. DisINR achieves better reconstruction quality than the evaluated INR baselines across multiple medical imaging tasks and retains performance advantages under the tested changes in image contrast, dataset, and sampling pattern. These results provide empirical support for the method’s effectiveness and transferability.
Critical weaknesses
  1. The paper does not sufficiently distinguish the contribution of the proposed architecture from the benefits of pretraining and restricting which parameters are updated during adaptation. The freezing ablation and experiments with controlled parameter counts support the overall design, but it remains unclear what additional benefits the explicit separation of shared and subject-specific encoders provides over a simpler approach based on pretraining and partial fine-tuning. Consequently, the contribution of the core architecture and the source of its performance gains remain insufficiently established.
  2. The tradeoff between preserving a fixed population prior and adapting to individual differences remains insufficiently characterized. Existing experiments show that freezing the shared modules is beneficial under the tested conditions. However, when the target subject’s anatomy, pathology, or acquisition conditions differ substantially from those represented in the pretraining data, it remains unclear whether the subject-specific encoder provides sufficient capacity for adaptation and when the shared representation itself needs to be updated.
Questions for authors
  1. Could the authors add a comparison with a pretrained INR that fine-tunes only a subset of its parameters, to assess the additional benefit of the DisINR architecture? Please use the same pretraining data and keep the trainable parameter counts and optimization budgets comparable.
  2. How much of the performance improvement depends on the shared encoder itself? Would a simplified model containing only a pretrained shared decoder and subject-specific encoders retain most of the gains, or does the explicit shared encoder provide a substantial additional benefit?
  3. Does the benefit of freezing the shared modules persist across the existing out-of-distribution evaluations? Are there particular changes in anatomy, pathology, or acquisition conditions under which partially or fully updating the shared modules becomes more effective than optimizing only the subject-specific encoder?
Reviewer_WJdqOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DisINR for medical image reconstruction by separating shared and subject-specific representations. A shared encoder–decoder pair and individual encoders are jointly pretrained using measurement-domain consistency. For an unseen case, the shared modules are frozen while a new subject-specific encoder is optimized. Evaluations on 3D volume fitting, undersampled MRI, and sparse-view CT report improvements in reconstruction quality and adaptation efficiency.

Critical strengths

++ The paper addresses an important limitation of scan-specific INRs: repeatedly optimizing an entire network without effectively reusing population information. Separating shared and individual learning provides a clear, practical approach to balancing prior reuse and case-specific adaptation.

++ Beyond reconstruction comparisons, the paper examines feature representations, parameter freezing, and alternative backbones. These analyses provide useful evidence for the proposed design, although they do not establish statistical disentanglement.

Critical weaknesses

-- "Seeking Common Ground While Reserving Differences: Multiple Anatomy Collaborative Framework for Undersampled MRI Reconstruction", IEEE IJBHI, 2026. already separates anatomy-shared and anatomy-specific learners. Although DisINR differs in its INR formulation, subject-level specialization, and test-time optimization, the conceptual relationship is not discussed. This omission makes the incremental contribution beyond existing shared–specific learning frameworks difficult to assess.

--Despite covering several datasets, individual experiments remain small. The 3D fitting experiment includes only three test volumes, while the KiTS23 and HNSCC stress tests each use five slices from a single volume. Independent patient counts and patient-level separation are also insufficiently specified for the main MRI/CT experiments. These limitations weaken confidence in the generalizability of the reported gains.

Questions for authors
  1. How does DisINR relate to previous shared–specific reconstruction framework, and which methodological advances go beyond applying this general principle to subject-specific INRs?

  2. How many independent patients are included in each split, and are training and test sets strictly patient-disjoint? Are repeated-split or seed results available to establish the stability of the reported improvements?

Comments on the paper

2 comments
Reviewer_uqgROpenReview ↗

Responses to Authors

Thank you for the additional experiments. The rebuttal substantially alleviates my original concerns. I will maintain my weak accept recommendation.

Thank you very much for your follow-up and for taking the time to consider our additional experiments. We sincerely appreciate your thoughtful evaluation and are grateful for your continued support of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.