Calibrated per-pixel uncertainty for low-dose paediatric chest-radiograph denoising at no fidelity cost
- 1University of Oxford
Abstract
BackgroundDeep denoisers restore low-dose chest radiographs but emit a single point estimate with no indication of where the output is reliable, so a confident-looking reconstruction can be locally wrong. In paediatric radiography, where the radiation-dose imperative is sharpest, we ask whether a per-pixel uncertainty map can be attached at no meaningful cost to reconstruction quality.
MethodsWe develop PP-VAE-Hformer, a hybrid convolution-transformer denoiser with a variational-autoencoder bottleneck (epistemic uncertainty map) and a heteroscedastic dual head (aleatoric map), and run a 19-arm ablation isolating five composite-loss terms on Poisson-Gaussian-degraded paediatric chest radiographs (Kermany collection; 624-image held-out test set), against eight discriminative baselines retrained on identical data and noise. Calibration is assessed by reliability diagrams and post-hoc{sigma} -scaling; comparisons use Bonferroni-corrected Welch tests with Cohens d as the primary discriminant.
ResultsA heteroscedastic negative-log-likelihood (NLL) objective buys a calibrated aleatoric map for 1.2 dB PSNR, which two-stage fine-tuning recovers to a statistically indistinguishable 0.010 dB. For the variational variant the Kullback-Leibler (KL) annealing schedule, not the bottleneck itself, is decisive: cyclic annealing closes a 1.10 dB gap over posterior-collapsing linear warmup. Predicted uncertainty tracks realised error (Pearson r > 0.8) with an optimal recalibration scale within 1% of unity. A matched-severity cross-degradation test shows this calibration is model-circular, however: at equal severity the aleatoric map becomes modestly (8-16%, a lower bound) over-confident once the noise model changes, and the Monte-Carlo epistemic map is a weaker error predictor that does not compensate. We additionally document a previously unreported destructive interaction between structural-similarity and Sobel-edge supervision at literature-default weights.
ConclusionsA calibrated per-pixel uncertainty map can be attached to a paediatric low-dose CXR denoiser at effectively no reconstruction cost. Its near-perfect calibration is tied to the trained ("inverse-crime") noise model, and the Monte-Carlo epistemic map does not reliably substitute where the aleatoric map fails, so an architecture-independent uncertainty estimate (e.g. deep ensembling), a blinded reader study, and real low-dose data are the immediate next steps.
Related
- Calibration, Uncertainty, and Data Efficiency of Deep Learning Chest Radiograph Classifiers Under Cross-Dataset Distribution Shift: A Multi-Dataset Benchmark
- CT-Referenced Diagnostic Performance and Complementary Error Patterns of Artificial Intelligence Versus a Single Emergency Physician in Chest Radiograph Interpretation: A Retrospective Paired Diagnostic Accuracy Study
- A Case of Asymptomatic Carotid Artery Stenosis With Thumb‐Sized Calcification Detected on Panoramic Radiograph
- When Does Data Augmentation Help with Few Chest-Radiograph Training Patients?
- Phantom Fairness: A Reproducible Audit of How Automatically Extracted Labels Can Conceal Demographic Disparities in Chest Radiograph Classifiers
BibTeX
@misc{mengistu2026calibrated,
title = {{Calibrated per-pixel uncertainty for low-dose paediatric chest-radiograph denoising at no fidelity cost}},
author = {S. T. Mengistu and K. Flouris},
howpublished = {medRxiv},
year = {2026},
month = sep,
doi = {10.64898/2026.09.14.26363026},
url = {https://www.medrxiv.org/content/10.64898/2026.09.14.26363026v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Calibrated per-pixel uncertainty for low-dose paediatric chest-radiograph denoising at no fidelity cost
AU - S. T. Mengistu
AU - K. Flouris
PY - 2026
DA - 2026/09/14
PB - medRxiv
DO - 10.64898/2026.09.14.26363026
UR - https://www.medrxiv.org/content/10.64898/2026.09.14.26363026v1
AB - BackgroundDeep denoisers restore low-dose chest radiographs but emit a single point estimate with no indication of where the output is reliable, so a confident-looking reconstruction can be locally wrong. In paediatric radiography, where the radiation-dose imperative is sharpest, we ask whether a per-pixel uncertainty map can be attached at no meaningful cost to reconstruction quality. MethodsWe develop PP-VAE-Hformer, a hybrid convolution-transformer denoiser with a variational-autoencoder bottleneck (epistemic uncertainty map) and a heteroscedastic dual head (aleatoric map), and run a 19-arm ablation isolating five composite-loss terms on Poisson-Gaussian-degraded paediatric chest radiographs (Kermany collection; 624-image held-out test set), against eight discriminative baselines retrained on identical data and noise. Calibration is assessed by reliability diagrams and post-hoc{sigma} -scaling; comparisons use Bonferroni-corrected Welch tests with Cohens d as the primary discriminant. ResultsA heteroscedastic negative-log-likelihood (NLL) objective buys a calibrated aleatoric map for 1.2 dB PSNR, which two-stage fine-tuning recovers to a statistically indistinguishable 0.010 dB. For the variational variant the Kullback-Leibler (KL) annealing schedule, not the bottleneck itself, is decisive: cyclic annealing closes a 1.10 dB gap over posterior-collapsing linear warmup. Predicted uncertainty tracks realised error (Pearson r > 0.8) with an optimal recalibration scale within 1% of unity. A matched-severity cross-degradation test shows this calibration is model-circular, however: at equal severity the aleatoric map becomes modestly (8-16%, a lower bound) over-confident once the noise model changes, and the Monte-Carlo epistemic map is a weaker error predictor that does not compensate. We additionally document a previously unreported destructive interaction between structural-similarity and Sobel-edge supervision at literature-default weights. ConclusionsA calibrated per-pixel uncertainty map can be attached to a paediatric low-dose CXR denoiser at effectively no reconstruction cost. Its near-perfect calibration is tied to the trained ("inverse-crime") noise model, and the Monte-Carlo epistemic map does not reliably substitute where the aleatoric map fails, so an architecture-independent uncertainty estimate (e.g. deep ensembling), a blinded reader study, and real low-dose data are the immediate next steps.
ER -