eess.IVNov 5, 2026

Measurement-Aware Multimodal LLM for No-Reference Low-Dose CT Quality Assessment

Authors: Kerem Basim, Tolga Demiroglu, Mehmet Ozan Unal, Metin Ertas, Isa Yildirim

Organizations: Dept. of Electronics and Communication Eng., Istanbul Technical University, Istanbul, Turkey · Dept. of Electrical and Electronics Eng., Istanbul University, Istanbul, Turkey

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

No-reference image quality assessment (IQA) of low-dose CT predicts the score that radiologists would assign to a slice without a full-dose reference, as needed to compare dose protocols and denoisers. Trained CT IQA networks agree well with radiologists but output only a score, whereas multimodal large language models (MLLMs) can explain a score but perform poorly on CT when only prompted. We propose a measurement-aware MLLM, built on Qwen3.8-27B, that rates a slice from its image together with six image-domain measurements of noise, sharpness, high-frequency content and streaks, given to the language model as percentiles in text. On the LDCT-IQA 2023 test set, it reaches an Overall score of 2.735, on par with the retrained challenge winner (2.735) and MedSigLIP-FiLM (2.713), and substantially above its base MLLM used as a zero-shot agent. Shuffling the measurements at test time reduces the score by 0.048, mostly through noise, but the same measurements also lead the model to assign higher scores to over-smoothed denoised images.

Figures & tables

Fig. 1: Two ways to give Qwen3.8-27B CT quality measurements. (a) Zero-shot agent: it reads precomputed tool outputs (measurements and visual views) in a loop, 2.8 reads per image on average, and writes a score and a rationale (Overall 2.18). (b) Our fine-tuned model: six measurement percentiles are written as text after the image tokens of a CT-tuned vision encoder; LoRA adapters and a linear head on the last token are trained to predict the MOS (Overall 2.73). Tiles are from a training slice and computed with simple definitions for illustration.
MethodPLCC ↑\uparrowSROCC ↑\uparrowKROCC ↑\uparrowOverall ↑\uparrowOverall, grid ↑\uparrowΔ\Delta [95% CI]τˉ\bar{\tau} ↑\uparrow
Prompted MLLMs
LMM-IQA, o3 [ 5 ].8122.8180.63472.2649---
Qwen3.8-27B agent (ours, Z3).7825.7859.61172.18022.1919−- 0.533 [ −- 0.668, −- 0.415]-
Fine-tuned MLLMs
MedGemma-4B QLoRA, own split [ 7 ].791.795-----
Q-Align [ 8 ].9095.9055.74222.5572 †2.5779−- 0.156 [ −- 0.210, −- 0.109]0.392
TABLE I: Results on the LDCT-IQA 2023 test set (300 slices) and on denoised images ( τˉ\bar{\tau} ). All models we ran are trained on our folds as 5-fold ensembles and evaluated under a protocol fixed in advance (our model’s selection: Sec. II ); LMM-IQA and MedGemma are cited from their papers. Grid: predictions rounded to the 0.2 MOS grid. ±\pm : std over three training seeds (point values are the original runs; for MedSigLIP-FiLM its best seed). Δ\Delta : difference to MedSigLIP-FiLM, paired bootstrap over the 300 slices. † Holm-adjusted p<0.05p<0.05 against ours. Bold: best of the rows above the bottom block. Bottom block: published results and our configurations chosen on the test set, not comparable to the rows above.
Fig. 2: What the models rely on. (a) Our model, partial dependence (fold 1): its mean prediction moves mostly with the noise measurements. (b) The zero-shot agent opens a visual view most often for the worst images (MOS 0–1); numbers on the bars are reads per image. (c) Agent absolute error by the measurement group its rationale cites (95% bootstrap CIs).
 ConfigurationOverall ↑\uparrowΔ\Delta [95% CI]τˉ\bar{\tau} ↑\uparrow
Zero-shot Qwen3.8-27B (no training)
Z1Image only1.370--
Z2+ meas. as text2.157+0.787 [+0.545, +1.028] †-
Z3+ agent loop (files, views)2.180+0.023 [ −- 0.125, +0.173]-
Z4+ 80 examples with MOS2.483+0.302 [+0.179, +0.436] †-
Trained, no LLM
TABLE II: Ablation. Each row changes one component, with the same folds, seed and 5-fold ensemble. Meas.: the six measurement percentiles as text. Δ\Delta : against the row above. † Holm-adjusted p<0.05p<0.05 . Bold: our model.
 OursZero-shot agent (%)
GroupΔ\Delta shuffled [95% CI]readcitedagrees [95% CI]
Noise level−- 0.024 [ −- 0.043, −- 0.007] †10099.7 ∗76.8 [72.0, 81.9]
Sharpness−- 0.005 [ −- 0.012, +0.001]10052.390.3 [85.7, 94.2]
HF residual−- 0.002 [ −- 0.007, +0.002]10010.387.1 [74.2, 96.8]
Streaks−- 0.006 [ −- 0.014, +0.001]10015.0 ‡80.0 [68.9, 91.1] ‡
Visual views-66.0--
TABLE III: What the models rely on. Ours: change in Overall when one measurement group is shuffled across test images (5 permutations). Agent: share of images on which it read each group, cited it (frozen keyword rules), and the citation agreed with the measured percentile. † Holm-adjusted p<0.05p<0.05 . ∗ A human reader coded noise in 90 to 94%. ‡ Rules not confirmed by a human reader ( κ<0.7\kappa<0.7 ).
More in this topic

Medical Image Quality Assessment

See all 26 papers

Explore similar work

CardsList

  1. Can MLLMs Perceive and Describe Medical Image Quality Beyond Scores?

    Oct 2, 2025Jiyao Liu, Jinjie Wei, Wanying Qu +16Medical Image Quality AssessmentMedical Image Benchmarks

    Fudan University · Shanghai Artificial Intelligence Laboratory · University of Cambridge

  2. MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment

    Aug 19, 2026Yuan li, Youyuan Lin, Chenhui Chu +1Image Quality AssessmentFaithfulness in VLMs

    Graduate School of Informatics, Kyoto University

  3. IQA-T1: Tool-based Visual Evidence Reasoning for Image Quality Assessment

    Jul 14, 2026Jinjian Wu, Jiaqi Tang, Wei Wei +5Tool Use in VLMsVLM Reasoning

    School of Computer Science, Northwestern Polytechnical University, Xi’an, China · The Hong Kong University of Science and Technology, Hong Kong SAR, China