Machine-Centric Image Quality Assessment

• MIQA is a task-driven framework that assesses image quality based on its effect on downstream machine vision tasks rather than human perception.
• Multiple formalizations—including machine preference tuples, ensemble accuracy, regression models, and causal feature learning—define image quality relative to task success.
• MIQA methods are applied in both general and medical imaging pipelines to enhance dataset auditing, quality control, and performance robustness.

Machine-Centric Image Quality Assessment (MIQA) is a task-driven, machine-oriented alternative to traditional Human Visual System (HVS)-based image quality assessment. Instead of estimating how good an image looks to a human observer, MIQA quantifies how image degradations affect the performance, stability, or task amenability of a downstream machine vision system or medical image analysis pipeline. Recent work has operationalized this idea through machine preference databases, consistency- and accuracy-based labels, causal feature learning, task-coupled controllers, and modality-specific quality assurance systems for MRI, CT, ophthalmic imaging, and vision-LLMs (VLMs) in medical settings (Wang et al., 27 Aug 2025, Li et al., 13 Mar 2025, Saeed et al., 2022, Ouaari et al., 2 Jul 2026).

1. Conceptual shift from human preference to machine utility

The central distinction in MIQA is that quality is defined relative to downstream algorithmic performance rather than perceptual fidelity. In this formulation, the relevant question is not whether an image is visually appealing or diagnostically comfortable for a reader, but whether the image preserves the visual evidence required for classification, detection, segmentation, retrieval, captioning, or clinical automation. This view is stated explicitly in work that defines MIQA as a task-driven, machine-oriented alternative to HVS-based IQA and in work that formalizes “machine subjective preference” through downstream tasks, test models, and evaluation metrics (Wang et al., 27 Aug 2025, Li et al., 13 Mar 2025).

This shift is motivated by systematic mismatches between human and machine sensitivities. Traditional full-reference and no-reference IQA metrics such as PSNR, SSIM, LPIPS, DISTS, HyperIQA, and MANIQA were developed to correlate with human perceptual judgments, not with machine task utility. In the MIQA literature, distortions in regions of interest, imperceptible yet harmful perturbations, and task granularity differences are treated as primary causes of this mismatch. Detection and instance segmentation, for example, depend strongly on spatial detail and localization, whereas classification often tolerates a different degradation profile (Wang et al., 27 Aug 2025).

The same distinction appears in medical imaging, where machine-centric MIQA focuses on whether an image is suitable for automated or assisted workflows. In this setting, quality is tied to whether clinically relevant structure is preserved for downstream AI, not simply whether a clinician can interpret the image. A zero-shot benchmark of 16 VLMs on medical image quality further showed that current models are sensitive not only to corruption patterns but also to textual context, indicating that machine-centric MIQA must account for representation stability and metadata robustness as well as image content (Ouaari et al., 2 Jul 2026).

2. Formalizations of MIQA targets

Several mathematically distinct, but compatible, formalizations of MIQA have emerged. One line of work defines machine preference through a tuple $(T, M, E)$ consisting of downstream task, test model, and evaluation metric. In that setting, pairwise machine preference can be induced from differences in downstream performance between a reference image and a distorted image, while aggregated machine Mean Opinion Score (MOS) is constructed from normalized scores across dimensions such as Yes-or-No, Multiple-Choice Questions, VQA, CAP, and a combined “Others” dimension for segmentation, detection, and retrieval (Li et al., 13 Mar 2025).

A second formalization defines MIQA labels through ensemble-aggregated consistency and accuracy. For a degraded image $x'$, consistency compares model outputs on $x'$ to outputs on the pristine image $x$, while accuracy compares outputs on $x'$ to ground truth. These quantities are aggregated into a mean machine opinion score:

$$E_C(x,x')=\sum_i \alpha_i C_{m_i}(x,x'), \qquad E_A(x,x')=\sum_i \alpha_i A_{m_i}(x,x'),$$

with optional composite score

$S(x',T)=\lambda_1 E_C(x,x')+\lambda_2 E_A(x,x').$

This formulation is explicitly used in the end-to-end MIQA paradigm and underlies benchmark labels for classification, detection, and instance segmentation (Wang et al., 27 Aug 2025).

A third formulation treats quality as task amenability. In that framework, a controller $h(x;\theta)\in[0,1]$ predicts the amenability of an image for a downstream task predictor $f(x;w)$, and the task predictor is trained with a weighted loss

$$\min_w \; \mathbb{E}\big[L_f(f(x;w),y)\,h(x;\theta)\big].$$

The controller is optimized so that images it scores as more amenable improve validation performance, yielding a bi-level objective in which MIQA is directly tied to downstream task success rather than to human labels (Saeed et al., 2022).

A fourth formulation introduces causal language. In Probability of Necessity and Sufficiency (PNS)-based MIQA, the goal is to learn hidden features that are necessary and sufficient for a high-quality outcome. Under exogeneity and monotonicity, the paper adapts PNS to learned hidden features as

$x'$0

and uses this quantity to separate causal features for “Good” images from complementary features associated with “not-Good” outcomes (Chen et al., 2024).

Taken together, these formalizations show that MIQA is not a single metric family but a broader design space in which image quality is anchored to machine behavior: prediction agreement, absolute task accuracy, sample usefulness for learning, or causal sufficiency for a target outcome.

3. Datasets and benchmark infrastructures

Large-scale MIQA required new datasets because human-opinion datasets do not encode downstream machine preference. The resulting resources differ in supervision granularity, task coverage, and whether they are full-reference, no-reference, or application-specific.

Resource	Scope	Labels or target
MPD	1,000 reference images, 30 corruption types, 5 levels, 30,000 reference/distorted pairs, 75 machine subjects, about 2.25 million fine-grained annotations	Machine MOS and per-dimension scores for YoN, MCQ, VQA, CAP, and Others (Li et al., 13 Mar 2025)
MIQD-2.5M	2.5 million degraded samples from 10,000 pristine images, 75 vision models, 250 degradation settings, 3 tasks	Consistency, accuracy, and composite MMOS for classification, detection, and instance segmentation (Wang et al., 27 Aug 2025)
PCMP	2,000 COCO references, 66 distortions per image, 327,241 PSNR-matched labeled pairs	Pairwise predictive-consistency preferences from seven pretrained models (Ding et al., 10 May 2026)
AS-OCT dataset	2,825 images: Good 593, Limited 1,827, Poor 405	Good vs. Deficient, where Deficient aggregates Limited and Poor (Chen et al., 2024)
MediMeta-C	Seven modalities, seven corruption types, five severity levels, 16 open-source VLMs	Zero-shot 1–5 quality ratings under corruption and bias prompts (Ouaari et al., 2 Jul 2026)
MRI cohort	250 patients	HO-derived 5-point Likert labels for reference-free MRI quality assessment (Küstner et al., 2018)
Clinical CT cohorts	NLST with 17,392 scans and internal datasets with 5,274 scans	Rule-based objective QA flags and batched subjective review for machine-learning readiness (Gao et al., 2021)

These resources made it possible to benchmark MIQA across substantially different settings. MPD was introduced as a machine preference database spanning natural scene images, screen content images, and AI-generated images, with machine subjects that include both large multimodal models and conventional computer-vision models (Li et al., 13 Mar 2025). MIQD-2.5M emphasized breadth of degradations and model diversity, explicitly covering region-dependent distortions through uniform distortion, ROI-dominated distortion, and background-dominated distortion (Wang et al., 27 Aug 2025). PCMP focused on PSNR-matched pairs so that preference labels reflect predictive consistency rather than simple distortion magnitude (Ding et al., 10 May 2026).

Medical resources serve a different role. The AS-OCT dataset supports causal MIQA under out-of-distribution scenarios (Chen et al., 2024), MediMeta-C exposes corruption and metadata sensitivity in medical VLMs (Ouaari et al., 2 Jul 2026), the MRI cohort supports reference-free quality classification from human-observer-derived labels (Küstner et al., 2018), and the CT pipeline datasets operationalize MIQA as structural and geometric fitness-for-use in large clinical repositories (Gao et al., 2021).

4. Model families and learning strategies

One major family of MIQA methods learns quality jointly with a downstream task. In the meta-reinforcement learning framework for image quality assessment, an IQA controller predicts per-image task amenability and uses those scores to weight or select samples for a task predictor. The controller is trained with reinforcement learning through validation rewards, with DDPG used in single-environment experiments and PPO in the meta-RL setting. The same work showed that task-specific, adaptable MIQA improved prostate presence classification, prostate gland segmentation, and pneumonia detection relative to non-selective baselines, and that meta-RL achieved comparable performance with only 20–30% expert labels for adaptation in both tasks (Saeed et al., 2022).

A second family aims at robustness through causal feature separation. The PNS-based MIQA framework for AS-OCT uses a feature extractor $x'$1, a complementary extractor $x'$2 with the same architecture, and a shared predictor $x'$3 with hidden layers $x'$4. The training objective

$x'$5

combines a standard prediction loss, a complementary loss that forces complement features toward “not-Good,” and a monotonicity term designed to stabilize the latent necessity-and-sufficiency interpretation. Across VGG16, ResNet50, EfficientNet-B0, and RegNet50 backbones, the framework improved in-distribution F1 and yielded consistent gains in the harder limited-only out-of-distribution split (Chen et al., 2024).

A third family treats MIQA as direct score regression over machine-generated labels. RA-MIQA is a region-aware transformer in which degraded images are split into patch embeddings, a frozen pretrained LUSS network supplies a region token, and a transformer encoder fuses region, class, and patch tokens before regressing a scalar quality score. The model is trained with mean squared error against consistency, accuracy, or composite labels. On MIQD-2.5M, RA-MIQA outperformed seven HVS-based IQA metrics and five retrained classical backbones, with reported SRCC gains of 13.56% on consistency and 13.37% on accuracy for image classification, and especially large gains in ROI-dominated and background-dominated settings (Wang et al., 27 Aug 2025).

A fourth family remains full-reference but replaces human perceptual supervision with predictive-consistency supervision. ML-CLIPSim defines machine-oriented quality as latent machine utility approximated through pairwise model votes on PSNR-matched distortion pairs. It uses a frozen CLIP ViT-B/16 visual encoder, aggregates intermediate patch-token similarities and global image embeddings, and trains only lightweight aggregation parameters with a pairwise binary cross-entropy loss on soft preference labels. On the PCMP test split, ML-CLIPSim surpassed PSNR, MS-SSIM, DISTS, LPIPS, and CLIPScore in ACC, SRCC, KRCC, and PLCC, and the same metric improved rate–task trade-offs when used as the distortion term in learned image compression (Ding et al., 10 May 2026).

These model families define the current methodological core of MIQA: task-coupled selection, causal representation learning, region-aware regression, and machine-preference full-reference similarity. They differ in supervision source, deployment assumptions, and whether the target is blind quality prediction or reference-based utility estimation, but all define quality through its effect on machine behavior.

5. Medical imaging implementations

In medical imaging, MIQA has taken both reference-free and workflow-integrated forms. A reference-free MRI framework trained on human-observer-derived labels was proposed to assess image quality immediately after each acquisition. The framework includes active learning, a blinded reading platform to reduce labeling effort, and classifiers based on support-vector-machine and deep neural network models. On a cohort of 250 patients, it achieved a test accuracy of 93.7% for estimating quality classes on a 5-point Likert scale, and was positioned for prospective quality assurance, automatic acquisition adaptation, guided MR scanner operation, retrospective quality assessment, and quality control in cohort studies (Küstner et al., 2018).

Clinical CT work has emphasized structural integrity and geometric consistency for machine learning pipelines. The CT quality assessment tool performs objective QA at the DICOM and NIfTI levels, including Instance Number validation, slice distance consistency checks, filtering of scans with few slices, physical length filtering, orientation checks, and resolution filtering. It uses dcm2niix for conversion, fslreorient2std for reorientation, optional lung ROI cropping, and FSL slicesdir for scalable batched human review. Applied to the National Lung Screening Trial, the pipeline identified approximately 4% of image volumes with quality concerns out of 17,392 scans and was then shown to generalize to internal datasets (Gao et al., 2021).

Ophthalmic imaging provides a more explicitly machine-centric label design. In the AS-OCT setting, the task is reduced to Good versus Deficient, where Good images have a complete, clear anterior chamber view and Deficient images include Limited and Poor cases, such as cropped chambers or significant artifacts that block analysis. The motivation is not merely perceptual quality but whether automated analysis can proceed reliably; the paper states that features necessary and sufficient for Good should be emphasized over merely correlated cues such as eyelashes (Chen et al., 2024).

A more recent medical direction uses VLMs as zero-shot quality raters. On the MediMeta-C benchmark, 16 open-source VLMs were evaluated across seven corruption types and five severity levels over seven modalities. Pixelation produced the largest score reductions, with mean $x'$6 and up to $x'$7 for OCT, whereas brightness had limited effect at about $x'$8. The same benchmark showed that textual attributes altered quality scores: institutional prestige raised them by $x'$9, equipment age lowered them by $x'$0, and the largest observed changes were $x'$1 for InternVL-8B and $x'$2 for MedGemma. These results establish that medical MIQA with current VLMs is affected by both corruption sensitivity and contextual bias (Ouaari et al., 2 Jul 2026).

6. Limitations, controversies, and emerging directions

A recurrent finding is that HVS-based metrics are misaligned with machine needs. On MIQD-2.5M, HVS metrics were described as inadequate for MVS quality prediction, and even specialized MIQA models still struggled with background degradations, accuracy-oriented estimation, and subtle distortions. Background-dominated distortion was consistently the hardest spatial regime, and accuracy was harder to estimate than consistency because it entangles image quality with model shortcomings on pristine inputs (Wang et al., 27 Aug 2025).

Benchmark evidence also shows that subtle degradation remains an unresolved problem. On MPD, learning-based no-reference and full-reference methods correlated reasonably well with machine MOS under severe distortion, but under mild distortion no method consistently surpassed PSNR on PLCC. The same resource reported that machine subjects disagree more with each other than human subjects do, with mean SRCC 0.62 among machine subjects versus 0.76 among human subjects, and that inter-task correlation is low, especially for VQA and specialized CV tasks (Li et al., 13 Mar 2025).

Causal MIQA introduces its own assumptions. The PNS framework depends on exogeneity and monotonicity, and the complement feature extractor is learned rather than generated by an explicit structural causal model. The paper also notes that device-level domain shifts, cohort differences, and acquisition settings were not evaluated, so broader robustness claims remain open (Chen et al., 2024).

Medical VLM-based MIQA exposes a distinct controversy: the relation between privacy protection and reliability. In the MediMeta-C benchmark, pixelation, described as a privacy-preserving transformation, substantially reduced scores, especially for OCT, indicating a trade-off between patient privacy and reliable automated quality evaluation. The same study concluded that metadata is simultaneously a privacy risk and a bias source because textual attributes unrelated to visual content changed scores materially (Ouaari et al., 2 Jul 2026).

Dataset-centric work adds another layer to the picture. A study on image data quality in machine learning models found that not all quality issues exert the same level of impact: convolutional neural networks were resilient to certain distortions but particularly vulnerable to degradations that obscure critical visual features, such as blurring and severe downscaling. In that workflow, automatic thresholding improved F1 from 0.6794 to 0.9468 under single perturbations and from 0.7447 to 0.8557 under dual perturbations, while the proposed deduplication strategy increased F1 from 0.4576 to 0.7928. This suggests that MIQA increasingly includes not only score prediction but also dataset auditing, threshold calibration, and redundancy control (Chen et al., 29 Sep 2025).

The overall trajectory is therefore toward a broader machine-centric quality ecosystem rather than a single universal metric. Current evidence supports task-aware labels, region-aware modeling, machine-preference benchmarks, causal robustness objectives, workflow-integrated medical QA, and explicit control of metadata and dataset pathologies. A plausible implication is that future MIQA systems will remain application-conditioned: the quality definition appropriate for MRI acquisition adaptation, CT cohort curation, AS-OCT analysis, or large-scale detection benchmarks is unlikely to collapse into a single human-style notion of perceptual quality.