One Shared LoRA Weight for MRI Reconstruction across Acceleration Factors
Zhiwei Zhao, Weikang Gong, Zhongnian Li, Xinzheng Xu
cs.CV · cs.LG
Quick verdict
Single shared LoRA with factor-conditioned gating for multi-factor MRI recon; promising, but text-table number mismatches and undiscussed losses to plain LoRA.
Summary
The paper proposes Shared LoRA, a parameter-efficient framework for accelerated MRI reconstruction that supports multiple acceleration factors with a single set of adapter weights. The pretrained SHFormer backbone is fully frozen; LoCon-style low-rank convolutional branches (rank 8) are inserted into selected encoder/decoder convolutions, and a lightweight MLP (GateNet) maps the scalar acceleration factor to per-layer gate coefficients that scale each branch's residual. Training mixes randomly sampled acceleration factors and their Cartesian (IXI-T2) or Gaussian (fastMRI knee) masks, so one adapter set learns cross-factor knowledge. On IXI-T2 with 4x/8x/16x mixed training, Shared LoRA reaches 42.17 dB PSNR at 4x (+0.61 dB over the strongest baseline) and 34.01/29.94 dB at 8x/16x; on fastMRI knee it reaches 37.23 dB at 8x. Trainable adapters are about 5.3% of backbone parameters, and performance at unseen factors 3x/5x/10x stays stable. However, the narrative numbers in Section 4.3.1 contradict Tables 1-2 in several places (e.g., 34.05/29.93 dB and 31.51 dB appear only in the text), and in Table 3's only equal-configuration head-to-heads (both models trained on 4x/8x), plain LoRA beats Shared LoRA at unseen 10x (31.05 vs 30.62 dB) and 32x (24.47 vs 23.67 dB) - results the text never discusses. No statistical tests, no code release, and no full-finetuning or SOTA-reconstruction baselines are provided.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A single set of shared LoRA adapters plus a factor-conditioned GateNet covers multiple acceleration factors (4x-16x on IXI-T2, 8x-32x on fastMRI knee) with best-or-competitive PSNR/SSIM against LoRA/DoRA/MeLoRA trained identically.
- The trainable adapters plus GateNet amount to only about 5.3% of the backbone parameters (238,928 added on a 4,434,595-parameter SHFormer).
- Randomized mixed-factor training enables stable performance at unseen neighboring factors (3x, 5x, 10x), with GateNet interpolating gate values smoothly.
- Coarse single gating suffices: fine-grained per-branch gating adds at most 0.03 dB (Table 6).
Strengths
Real deployment problem, clean PEFT formalization
The per-factor model cost motivation (Section 1, citing robustness studies [8,15]) is genuine, and the method is precisely specified: LoCon-style rank-8 conv adapters (alpha=16, dropout 0.05) in selected encoder/decoder convolutions, GateNet MLP 1-64-64-D with unit-gate initialization so training starts exactly at the pretrained solution (Section 3.3, Eqs. 1-9).
Fair baseline design on identical mixed-factor training
In Tables 1-2, LoRA, DoRA and MeLoRA are all trained with the same randomized-factor input pipeline on the same frozen SHFormer backbone (Section 4.3.1), so the comparison isolates the adaptation mechanism rather than the training data.
Headline gains on primary configurations
Candid negative micro-ablation
Coherent parameter accounting and insertion-cost analysis
Weaknesses
Narrative numbers contradict the result tables
Undiscussed negative head-to-heads in Table 3
No ablation isolating GateNet
Narrow baseline scope and unmeasured cost claims
Weak statistical treatment and reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This method paper targets a documented pain point in deep MRI reconstruction: models are trained per acceleration factor, so supporting new factors multiplies training, storage, and deployment cost (Section 1, citing robustness studies [8,15]). The proposal is Shared LoRA: a frozen pretrained SHFormer backbone plus (i) one shared set of low-rank convolutional adapters and (ii) GateNet, an MLP conditioned on the scalar acceleration factor that emits per-layer gate coefficients (Section 3.3). Central claims: single-adapter coverage of 4x-16x (IXI-T2) and 8x-32x (fastMRI knee) with best-or-competitive PSNR/SSIM (Section 4.3.1, Tables 1-2); trainable parameters of about 5.3% of the backbone (Section 3.5); stable behavior at unseen 3x/5x/10x (Section 4.3.2, Table 3); and sufficiency of coarse gating (Section 4.3.4, Table 6). Scope is 2D, fully supervised, image-domain reconstruction with Cartesian (IXI) and Gaussian (fastMRI) masks.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Shared LoRA for MRI reconstruction?
It is a parameter-efficient fine-tuning method that freezes a pretrained SHFormer reconstruction network and trains one shared set of low-rank (LoRA-style) convolutional adapters plus a small gating network (GateNet) that takes the acceleration factor as input. The goal is to serve multiple acceleration factors (e.g., 4x-16x) with a single adapter set instead of maintaining one full model per factor.
How many extra parameters does Shared LoRA add?
According to the paper's insertion-location ablation, the adopted configuration adds 238,928 parameters (including GateNet) on a 4,434,595-parameter backbone - about 5.3-5.4% of the model - using rank-8 low-rank convolutional branches.
Does it generalize to acceleration factors not seen during training?
The paper reports stable results at unseen 3x, 5x and 10x after training on 4x/8x/16x, with 5x and 10x following the expected interpolation trend. One caveat from the paper's own Table 3: in the only equal-condition comparison (both models trained on 4x/8x), plain LoRA actually scored higher than Shared LoRA at unseen 10x and 32x, a result the paper does not discuss.
Is code available for Shared LoRA?
No code, checkpoint, or data release is mentioned in the paper. Experiments use public IXI-T2 brain MRI and a Hugging Face mirror subset of fastMRI knee, but training details such as learning rate, mask generation parameters, and slice extraction are not fully specified.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-3 are extracted as garbled text runs (architecture and qualitative-comparison labels), so visual details of Figure 2 could only be cross-checked against the prose, not verified pictorially; all numbers cited in this review were verified against the extracted Tables 1-6, and the body text shows no truncation (it ends with the reference list).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.