List of Papers Browse by Subject Areas Author List
Abstract
Adapting large foundation models to medical imaging is costly due to limited curated data, dense expert annotation requirements, and the expense of full fine-tuning. Parameter-efficient fine-tuning (PEFT) methods such as LoRA mitigate these costs, but linear low-rank updates can be expressivity-limited. Recent sinusoidal variants such as SineLoRA improve capacity, yet depend on a fixed frequency hyperparameter that is difficult to transfer across modalities and tasks. We propose \textbf{Prior-Conditioned Low-Rank Adapters (PriLoRA)}, an input-conditioned sinusoidal adapter that leverages input-derived prior features to modulate the adapter response per block and per sample while preserving LoRA’s modularity and parameter efficiency. PriLoRA predicts lightweight modulation parameters that control the sinusoidal update, enabling adaptive expressivity without increasing the adapter rank or requiring task-specific frequency search. We evaluate PriLoRA across multiple medical backbones, modalities, and tasks, including generation, classification, segmentation, and CT reconstruction, and observe consistent improvements in robustness and transfer over standard LoRA and representative PEFT baselines across datasets and modalities.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5999_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/xmindflow/PriLoRA
Link to the Dataset(s)
IU X-RAY Dataset: https://data.lhncbc.nlm.nih.gov/public/Tuberculosis-Chest-X-ray-Datasets/Shenzhen-Hospital-CXR-Set/Annotations/index.html
ISIC 2016: https://www.kaggle.com/datasets/mahmudulhasantasin/isic-2016-original-dataset
OCTMNISTv2: https://medmnist.com/
CT Reconstruction: https://www.kaggle.com/datasets/younesselbrag/lung-nodules-detection-dataset-annotations
BibTex
@InProceedings{KazAmi_PriLoRA_MICCAI2026,
author = { Kazerouni, Amirhossein AND Karimijafarbigloo, Sanaz AND Aghayari, Ali AND Ghasemzadeh Khorzoughi, Ali AND Azad, Reza AND Samek, Wojciech AND Merhof, Dorit AND Brudno, Michael AND Taati, Babak},
title = { { PriLoRA: Prior-Conditioned Low-Rank Adapters for Medical Vision Models } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes Prior-Conditioned Low-Rank Adapters (PriLoRA), a parameter-efficient adaptation method for medical vision models. It extends LoRA with an input-conditioned sinusoidal low-rank weight update instead of a fixed low-rank linear update. Prior features derived from inputs allow to predict modulation parameters that control the adapter response for each sample and block. PriLoRA seeks to avoid the manual frequency tuning of SineLoRA while improving expressivity without increasing adapter rank. Experiments across several medical imaging tasks (generation, segmentation, classification, and CT reconstruction), backbones, datasets, and modalities show that PriLoRA can provide more consistent performance than standard LoRA and PEFT baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper is well-written and structured. The key concepts are described in enough detail to make them understandable.
- PriLoRA introduces a light modulator that leverages prior features extracted from the input and then predicts a small set of parameters that specialize the adapter for each block and each sample. Sinusoidal LoRA therefore becomes an input-conditioned adapter that adjusts its non-linear response per block and per sample, while preserving LoRA modularity and parameter efficiency.
- The paper is well motivated. Instead of using a fixed linear LoRA update, or a fixed-frequency sinusoidal update (SineLoRA), PriLoRA predicts sample- and block-specific modulation parameters from input-derived prior features. This avoids manual frequency selection across tasks and modalities. To avoid adapting for different modalities and tasks, PriLoRA provides an adapter response that depends on the input.
- The experimental evaluation on four different medical tasks (image generation, segmentation, classification, and CT reconstruction). PriLoRA is shown to improve robustness over standard LoRA and several representative PEFT baselines.
- PriLoRA only adds a small overhead in terms of trainable parameters and latency.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- PriLoRA directly extends from variants like SineLoRA. The authors for not provide muckh on related works, so it is difficult to put the novelty of their work into perspective among SOTA adaptation methods.
- The paper does not isolate the origins and extent of improvements: from the sinusoidal formulation or the input conditioning? - The paper validates experimentally over several tasks, but usually with only one dataset per task, so robustness is not validated over different institutions, scanners, etc. (with distribution shifts).
- The authors should provide an analysis of computational complexity for their methods versus SOTA methods. - The code is not made available, so the results in this paper may be difficult for a reader to reproduce.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
None.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
I would recommend weak acceptance because the paper proposes a well-motivated extension of LoRA that is cost-efficient and improves performance across several medical imaging tasks. The score is tempered by incremental novelty, limited robustness analysis across datasets/sites, and missing details on complexity and reproducibility.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes PriLoRA, a prior-conditioned sinusoidal low-rank adapter for parameter-efficient fine-tuning of medical vision models. The key idea is to replace the globally fixed sinusoidal transformation used in SineLoRA with an input-conditioned one: a lightweight modulator extracts input-derived prior features and predicts four coefficients (alpha, beta, gamma,delta) that control the amplitude, frequency scaling, phase, and offset of the sinusoidal low-rank update. In this way, the method aims to preserve the expressivity benefits of sinusoidal low-rank adaptation while avoiding manual tuning of a single fixed frequency across heterogeneous medical tasks and modalities. The paper further positions PriLoRA as a general PEFT mechanism and evaluates it across multiple medical backbones and tasks, including generation, segmentation, classification, and CT reconstruction.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper has two main strengths. First, the method is simple and easy to understand. PriLoRA is a clean extension of LoRA/SineLoRA, with an intuitive formulation based on a lightweight modulator that predicts amplitude, frequency scaling, phase, and offset for the sinusoidal update. Second, the empirical evaluation is fairly comprehensive. The method is tested across multiple medical vision tasks, backbones, and modalities, and the paper also includes ablation studies and efficiency analysis in addition to the main quantitative results.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The design of the PriLoRA modulator appears to be highly task-dependent, rather than a single unified and broadly generalizable design.
2.Figure 1(c) still retains omega_0, while the paper claims that PriLoRA avoids manual frequency tuning. In addition, the roles of omega_0 and beta appear potentially overlapping, since beta rescales the sinusoidal argument that already contains omega_0.I would encourage the authors to further clarify the rationale behind this design choice and explain whether omega_0 still needs to be manually selected in practice.
3.The paper states that rank 8 is used for classification, but the numbers reported in Table 1 match the rank-16 results in Table 2 rather than the rank-8 results. The authors should clarify which rank was actually used for the main classification comparison.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper has a certain degree of novelty, and it presents targeted experiments across different scenarios with reasonably good results. However, the method description is not sufficiently clear, the design appears somewhat redundant, and there are concerns about the consistency of the experimental results. Therefore, I can only give it a weak accept. Overall, I remain neutral on whether this paper should be accepted.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes PriLoRA, a prior-conditioned low-rank adaptation method for medical vision-language models. PriLoRA builds on SineLoRA, which introduced non-linear sinusoidal activations for learnable LoRA parameters. Specifically, it introduces additional parameters in the sinusoidal modulation step that are derived from the input data. The motivation behind the approach is that different medical tasks (or priors) would require varying gains, frequencies, phase, etc. during the sinusoidal modulation. The approach outperforms prior low-rank adaptation methods.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The idea of conditioning the effective LoRA rank through the additional introduced parameters in the sinusoidal modulation is new and interesting. 2.The approach achieves competitive performance. 3.The paper is well-written and easy to understand.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The paper lacks evidence to substantiate several motivating claims. For instance, the “Effect of Parameters” paragraph in Sec. 3.1 states that alpha controls the ability of PriLoRA to attenuate or amplify easy and difficult scenarios, beta captures fine-grained variations, gamma avoids saturation for certain modalities and delta allows domain-specific shifts. However, none of these claims have been validated.
2.Analysis of learned alpha, beta, gamma and delta: While ablations show that removing each of these parameters affects performance, the paper does not provide insights into the learned values of the parameters for different tasks.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The idea is interesting but some claims are unjustified. The paper would also benefit from additional analysis.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper introduces PriLoRA, a prior-conditioned low-rank adaptation method for medical vision models. It extends LoRA/SineLoRA by using input-derived prior features to modulate the sinusoidal low-rank update, aiming to improve expressivity while keeping the adaptation parameter-efficient. The paper received consistent positive reviews, with all three reviewers recommending Weak Accept. Reviewers appreciated the clear motivation, simple formulation, good writing, and broad evaluation across multiple medical imaging tasks, including generation, segmentation, classification, and CT reconstruction. The main concerns are about the limited analysis of the learned modulation parameters, some unclear design choices, and reproducibility details. Overall, I recommend acceptance. The contribution is somewhat incremental relative to LoRA/SineLoRA, but the method is well motivated, technically sound, and evaluated across a reasonably broad set of medical imaging scenarios. The remaining issues can likely be addressed in the final version.