Lung cancer on low-dose CT: what if we trained AI on "degraded" full-dose scans instead of denoised ones?

A team at Politecnico di Milano (Italy) tested an idea that flips the problem: for lung cancer screening with low-dose CT (LDCT), rather than denoising noisy images, degrade the abundant full-dose scans (SDCT) into fake low-dose images to train a radiomics nodule classifier. Comparing three degradation methods on the public LIDC-IDRI dataset, the authors show that a model trained on full dose collapses on real low dose (sensitivity 0.571), and that all three degradations restore an AUC above 0.84 — the best, a CycleGAN, reaching 0.861 with a sensitivity of 0.743. But ranking the three methods against each other remains statistically fragile.

The context

Lung cancer screening in heavy smokers relies on low-dose CT, or LDCT. The principle: sharply cut the X-ray dose to limit the risk of repeated irradiation, at the cost of noisier, lower-contrast images sometimes marked by artifacts. Large trials (NLST in the US, NELSON in Europe) showed such screening reduces mortality. What remained was to automate the reading — spotting and classifying lung nodules — with machine-learning models.

But these models hit a shortage: annotated LDCT datasets are very scarce. The literature has therefore mostly worked on denoising low dose to bring it closer to full dose. Two problems with that: denoising often requires paired SDCT/LDCT scans of the same patient (ethically impossible to acquire, since it would mean scanning the same person twice), and the "degrade then denoise" chain is heavy. The Milan team proposes the reverse: start from the plentiful, annotated full-dose images and degrade them into fake low-dose images to augment the training data — then check that a classifier trained this way works on real low dose. The task measured is a binary classification (malignant nodule or not), assessed mainly by the AUC (area under the ROC curve), a score between 0.5 and 1: 0.5 is chance, 1 is perfect separation between malignant and benign.

The method

The heart of the work is a comparison of three ways to make "fake LDCT" from real SDCT. Method 1: noise insertion in the sinogram domain (the scanner's raw projection data), adding Poisson and Gaussian noise to mimic quantum and electronic noise — a deterministic, lightweight approach. Method 2: replicating a validated physics model (Yu et al.) via Pix2Pix, a conditional generative adversarial network (GAN) trained on paired images; a GAN pits a generator, which fabricates images, against a discriminator, which tries to tell real from fake, until the fake becomes convincing. Method 3: a CycleGAN, a variant that learns to translate from one domain (full dose) to another (low dose) without pairs, from two sets of unpaired images — valuable here, since real SDCT/LDCT pairs do not exist.

The data come from two public sources. The main one, LIDC-IDRI, gathers 1,018 thoracic CT scans annotated by radiologists. The authors split off 695 full-dose patients (1,950 nodules), reserved for training after degradation, and 315 low-dose patients (675 nodules) reserved exclusively for validation (50%) and testing (50%). A nodule is labeled "malignant" when the average likelihood assigned by radiologists exceeds 4 out of 5 — that is, 225 malignant versus 1,725 benign in training, a strong imbalance. The second source (the Low-Dose CT Image and Projection dataset, 50 paired chest exams) serves only to train Pix2Pix. Preprocessing is uniform: conversion to Hounsfield units, windowing [−1200, 600], light Gaussian denoising, normalization.

On each degraded set, radiomics enters: PyRadiomics extracts 851 quantitative features (shape, intensity, texture, wavelets) compliant with the IBSI standard, reduced by several selectors (LASSO, MRMR, PCA, RFE, random forests and gradient boosting). The imbalance is corrected by augmenting the minority class (contour perturbations bringing the 225 malignant to 900) and undersampling the benign (to 900). Six classifiers are tested (AdaBoost, decision tree, k-nearest neighbors, logistic regression, random forest, XGBoost). Half of the real LDCT is used to choose the threshold, feature subset and best selector/classifier pair; the other half, never seen, is the final test, with 1,000 bootstrap resamples for confidence intervals and Friedman then Wilcoxon tests (Bonferroni-corrected) to compare methods. A key rigor point: no low-dose patient takes part in training the classifier.

The results

First lesson: without degradation, it breaks. A classifier trained on raw full dose behaves well in validation (AUC 0.863) but drops on real low dose: AUC 0.789 and, above all, a sensitivity that collapses to 0.571, for a specificity of 0.917. In other words, out of 100 truly suspicious nodules, this model misses about 43 — precisely the error a screening tool cannot afford.

Second lesson: all three degradations repair the essentials. Each restores a test AUC above 0.84: 0.844 (method 1, sensitivity 0.656), 0.859 (method 2, sensitivity 0.684) and 0.861 for the CycleGAN (method 3), which offers the best balance — balanced accuracy 0.800, sensitivity 0.743, specificity 0.858. In clinical translation, we move from about 43 suspicious nodules missed per 100 (non-degraded model) to about 26 per 100 (CycleGAN); on the false-positive side, a specificity of 0.86 leaves about 14 benign nodules per 100 wrongly flagged as suspicious — a sensitive point in screening, where excess false positives trigger unnecessary work-ups and interventions.

Third lesson, subtler, about image quality. Measured by distances between distributions (FID and KID, where lower is better), alignment with real low dose is clearly best for the CycleGAN (FID 0.17) versus methods 1 and 2 (FID 4.07 and 4.67) — which do worse than the non-degraded full-dose images (FID 1.15). The paradox is instructive: method 2 produces visually very "realistic" noise and streaks, but statistically farther from real low dose. Realism as seen by the eye is not the measured alignment of distributions. Finally, despite differently-looking images, the retained radiomic features overlap heavily (60 features common to the three methods), and only two are selected everywhere: the nodule's major-axis length and its sphericity — consistent with the clinical Lung-RADS criteria, which judge first on size and regularity.

What is good

A genuine domain-shift stress test. The protocol's strength is strict separation: the real low-dose scans serve only to validate and test, never to train the classifier, which rules out data leakage. The demonstration that the full-dose model collapses on low dose (sensitivity 0.571) is no detail: it establishes in black and white that dose is a domain boundary one cannot ignore.

An honest decoupling of realism and alignment. By reporting both visual appearance and distributional distances, the authors show that an image that "looks" low dose (method 2, marked streaks) can be farther from real low dose than a blurrier one (CycleGAN). This is a useful guardrail against the illusion that synthetic data convincing to the eye is necessarily good synthetic data — and steering the CycleGAN's training with those distances is coherent.

Careful reproducibility. Public data (LIDC-IDRI and the Low-Dose CT Image and Projection set, both on The Cancer Imaging Archive), code released on GitHub, IBSI-compliant radiomics, bootstrap confidence intervals and explicit statistical tests. The work can be rerun and debated — the basis of any constructive critique.

What is less good

The target is not biological ground truth. "Malignant" is defined here by the average likelihood estimated by radiologists (> 4 out of 5), not by histological confirmation. The model therefore learns to reproduce a radiological impression, not a proven diagnosis: a potential misleading metric, not to be read as detection of actual cancer. Add to this a strong class imbalance (about 11% malignant) that augmentation and undersampling correct artificially.

Still-local generalization. The real low-dose validation and test scans all come from LIDC-IDRI: the domain shift tested is partly internal to a single dataset, with no truly external cohort (the authors point to a future test on LUNA25). Moreover, Pix2Pix is trained on only 50 exams from a single set. This population bias — a single ecosystem of devices and protocols — leaves open the question of transfer to other centers.

A fragile ranking between methods. The Friedman and Wilcoxon tests come out significant, but the 95% confidence intervals overlap heavily (AUC 0.859 versus 0.861 for methods 2 and 3), and the authors themselves acknowledge that the size of the bootstrap resamples makes the tests hypersensitive to tiny differences. The robust conclusion is therefore not "CycleGAN wins" but "the three degradations are roughly equivalent and clearly beat the absence of degradation." The 0.743 sensitivity also stays modest and comes with the widest intervals. Finally, note that the CycleGAN "saw" the distribution of the real low-dose images (without labels) during training — standard practice in domain adaptation, but worth flagging.

What it changes

For the research community, the study reframes a recurring low-dose problem: rather than denoising rare images, one can degrade abundant full-dose images to build training data — and unpaired translation (CycleGAN) offers a pragmatic route where real pairs are ethically inaccessible. The methodological message is twofold: domain adaptation is necessary (the naive model collapses), and the quality of a synthetic set is judged on distributional alignment, not on apparent realism.

For clinicians and decision-makers, the warning is concrete: an AI validated on full-dose diagnostic scans can fail silently on low-dose screening scans. The dose regime is a domain difference in its own right; a tool meant for screening must be developed and validated under screening dose conditions. And specificity matters as much as sensitivity, because the Achilles' heel of LDCT screening remains the false positive and its train of unnecessary examinations.

For patients and the public, two simple ideas. First, "screening scan" and "diagnostic scan" do not produce the same images, and an algorithm that performs on one is not necessarily good on the other. Second, a synthetic image that looks more realistic is not automatically better for training an AI. At this stage — AUC around 0.86, sensitivity around 0.74, on a single dataset and without histological confirmation — this is promising research, not a tool ready for the clinic.

To go further

The preprint "A Comparative Analysis of CT Degradation for LDCT Nodule Classification using Radiomics" is available on arXiv (2605.12164), dated 13 May 2026, by Jiaying Liu, Valentina D. A. Corino, Anna Corti and Luca Mainardi (Politecnico di Milano; Corino also at the Cardiotech Lab of Centro Cardiologico Monzino IRCCS, Milan). The code is released on GitHub, and the data are public via The Cancer Imaging Archive (LIDC-IDRI; Low-Dose CT Image and Projection Data). The preprint consulted carries no declaration of funding or conflicts of interest. On the blind spots of generalization and training data in imaging, see our decryptages on mixing datasets that degrades screening AI, on deep radiomic signatures and interpretability, and on hidden subgroup biases in medical imaging.

Editorial transparency: French version written and signed by the Tatakoto editorial team based on a reading of the preprint. English, Spanish and Chinese translations produced with AI assistance and reviewed.