acceptodds
Withdrawn from ICLR 2027

Task-Aware Color Constancy for Chronic Kidney Disease Prediction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Smartphone-based kidney health prediction offers a low-cost alternative to laboratory testing, but its accuracy degrades significantly under unconstrained lighting conditions that distort the colorimetric signals of creatinine test strips. Standard color constancy methods, which optimize for perceptual visual quality, often fail to preserve the precise spectral cues required for clinical quantification. To address this, we propose a Task-Aware Color Constancy (TACC) framework that explicitly aligns illumination correction with downstream disease prediction. Our approach integrates a conditional generative adversarial network (TACC-GAN) with NeurSPR, a neural structured prediction module. Crucially, we introduce a differentiable histogram layer based on Kernel Density Estimation (KDE), which bridges the “non-differentiable bottleneck” of traditional regressors. This allows diagnostic losses to backpropagate directly to the generator, optimizing color correction for clinical accuracy rather than visual appeal. We evaluate on a demographically stratified dataset constructed from US population statistics. Among deployable (reference-free) methods, TACC achieves the best kidney-health classification performance in every lighting condition, raising the F1-score from 0.50 to 0.87 under extreme lighting variations while keeping creatinine-estimation error competitive with the strongest baseline. Notably, it approaches oracle calibration methods that require a white-light reference at test time, despite using no such reference, opening the door towards reliable point-of-care monitoring under real-world settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
color constancy, conditional generative adversarial network, kidney health prediction, neural structured prediction

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_rB4eOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes TACC, a task-aware color constancy framework for smartphone creatinine-strip kidney-disease detection: TACC-GAN maps heterogeneous-lighting images to a canonical white-light domain via a tripartite-mask conditional GAN, and NeurSPR uses a differentiable histogram with soft MoE regression to backpropagate diagnostic losses. All experiments rest on 2,016 images of 504 laboratory-prepared strips under 4 scripted lighting conditions, reporting the best deployable classification F1 per condition. Overall, the workload is minimal, the figures are poor, and the exposition is unclear—far below the ICLR standard; I recommend strong rejection.

Critical strengths
  • The task-aware motivation has practical value. Aligning the optimization objective of color correction with downstream diagnostic accuracy rather than perceptual quality, applied to low-cost kidney monitoring, is a meaningful entry point.
Critical weaknesses
  • The workload is minimal and far short of a top-tier conference paper: everything rests on a single small dataset of 504 laboratory strips, and the claimed demographic stratification is merely synthetic metadata rather than real patients. The method directly assembles existing components, and NeurSPR's F1 gain over the strongest classical baseline is only 0.01—closer to a workshop paper.

  • Poor figures and sloppy presentation: the key results figure, Figure 4, labels 8-bit "Pixel Value 0–240" axes as Lab channels, directly contradicting the text, and its subplot scales are incomparable. Table 2 contains the "0.0.8515" error, and the KDE-overlap metric announced in Section 4.3 is never reported numerically anywhere.

  • Confusing exposition with missing key details: the color space, bin count K, and bandwidth σ of the differentiable histogram are unspecified, and the loss weights in Eq. (4) contradict Section 4.4. The generator/discriminator architectures and mask acquisition are also undescribed, hindering understanding and reproduction.

Questions for authors
  • What is the concrete dataset-expansion plan—what target scale, and will it include multiple phone models and self-captured images from real patients? Will there be validation against laboratory serum creatinine assays rather than eGFR labels derived from synthetic metadata?

  • Will you fix the axis/color-space contradiction in Figure 4 and errors such as "0.0.8515" in Table 2? Please report the KDE-overlap values (by method × lighting condition) rather than passing over them with a single qualitative figure.

  • Please clarify the differentiable histogram implementation (color space, K, µ_k, σ, t_k) and resolve the loss-weight contradiction between Eq. (4) and Section 4.4. The generator/discriminator architectures, image resolution, and mask acquisition also need to be supplied.

Reviewer_zLwGOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper addresses smartphone-based kidney health screening from colorimetric creatinine test strips under uncontrolled lighting. It proposes Task-Aware Color Constancy (TACC), which has two parts. The first is a mask-guided conditional GAN that maps strip images to a canonical white-light domain. The second is NeurSPR, a differentiable regressor built on a KDE-based soft histogram and a mixture of experts. The differentiable histogram lets diagnostic losses backpropagate into the generator, so color correction is optimized for clinical accuracy rather than perceptual quality. Experiments across several lighting conditions show that TACC outperforms learned and statistical color-correction baselines and performs comparably to reference-based calibration.

Critical strengths
  • The motivation is clear. The differentiable histogram is a simple, sensible way to connect correction and diagnosis end to end.
  • The evaluation covers multiple lighting conditions, both color-correction and predictor baselines, and ablations that isolate the effect of the masks and of the diagnostic feedback.
Critical weaknesses
  • The validation is too narrow for the clinical claims. All data come from a small set of lab-prepared strips captured in a single setup. There is no external test set and no cross-device or cross-batch evaluation, so the claims about real-world point-of-care use are not supported.
  • The benefit over calibration is not shown. Reference-based calibration performs on par with TACC. The paper treats calibration as non-deployable, yet the strip carries its own white patch and color anchors. Per-image calibration from those regions would seem to need neither a paired reference nor any training, and this baseline is missing. Without variance estimates or significance tests, the margins over the other methods are also hard to judge.
Questions for authors

No questions

Reviewer_5hByOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Task-Aware Color Constancy (TACC) for smartphone-based creatinine-strip analysis under variable illumination. TACC-GAN maps images acquired under four lighting conditions to paired white-light references. A mask-weighted reconstruction objective emphasises the reaction zone, RGB anchors, and white reference, while adversarial and angular losses constrain the transformation. The downstream NeurSPR module replaces hard histogramming and tree-based selective partitioned regression with Gaussian-KDE histogram features and a soft mixture of regression experts, allowing diagnostic losses to propagate to the correction generator. Predicted creatinine, assigned age, sex, and race variables are then used to derive three eGFR-based status classes. TACC reports the highest weighted F1 among reference-free methods and includes component ablations, but the current evidence does not substantiate the paper's clinical CKD or real-world deployment claims.

Critical strengths
  1. The paper targets a practically important failure mode of smartphone colorimetric assays: illumination correction should preserve task-relevant biochemical colour rather than merely improve visual appearance.
  2. The differentiable KDE and soft mixture-of-experts formulation provides a coherent route for passing a diagnostic regression signal through histogram-based features to the correction network.
  3. The study compares learned, statistical, and reference-based correction methods under four lighting conditions and clearly separates deployable reference-free methods from oracle calibration methods.
Critical weaknesses
  1. The dataset is constructed from 504 laboratory-prepared creatinine strips rather than samples obtained from 504 patients. Each strip is assigned a unique identifier, and its four lighting-condition images are kept within the same split. This grouped split appropriately prevents leakage between repeated images of the same physical strip, but it should be described as a strip-level split rather than a patient-wise split. Moreover, age, sex, and race are sampled from census statistics rather than obtained from actual participants, making the resulting eGFR and kidney-status labels partially synthetic. The experiments therefore demonstrate generalisation across held-out strips and lighting conditions, but do not establish demographic diversity, patient-level clinical generalisation, or CKD diagnostic performance.
  2. The claimed end-to-end diagnostic pathway is underspecified. Section 3.3 obtains status by hard thresholding eGFR, which is non-differentiable, whereas Eq. 10 assumes predicted class probabilities without defining the status head or its connection to creatinine. It is therefore unclear how the status loss reaches TACC-GAN.
  3. The optimisation description is internally inconsistent: NeurSPR is said to be pretrained on white-light images and frozen during TACC-GAN training, yet Section 4.4 says both modules are trained end-to-end. A frozen differentiable critic can transmit gradients, but that differs materially from joint end-to-end training.
  4. Evaluation uses a single dataset, camera/application pipeline, strip design, and laboratory acquisition protocol. Four lighting labels do not test new phones, camera ISPs, strip batches, environments, biological specimens, users, or sites, so the stated real-world robustness and point-of-care deployment claims are unsupported.
Questions for authors
  1. Are the 504 samples actual patient specimens or laboratory-prepared concentration strips? If they are laboratory prepared, why are strip IDs described as patient IDs and synthetic demographics described as demographic diversity?
  2. How are the predicted status probabilities in Eq. 10 computed? Please give the complete differentiable path from NeurSPR outputs through eGFR and status prediction to the generator.
  3. Is NeurSPR frozen or jointly updated during TACC-GAN optimisation? Please provide the training schedule, gradient-flow graph, and parameter groups updated by each loss.
  4. Were DeepWB and AngularGAN retrained on exactly the same paired training data with comparable tuning and compute? Did every correction method use the same crop, masks where applicable, and fixed NeurSPR evaluator?
  5. Can the authors report results over repeated strip-disjoint splits or cross-validation with confidence intervals and paired tests, including per-class sensitivity, specificity, macro-F1, calibration, and confusion matrices?
  6. Can the authors provide quantitative Lab/KDE-overlap results over the complete test set and validate on unseen phones, camera pipelines, strip batches, users, environments, and biological samples?

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

Related papers

Discussion (0)

Sign in to comment.