Do Low-Dose CT Denoisers Erase Rare Lesions? Phenotype-Balanced Training for Restoration

Deep networks that clean up low-dose CT scans learn from training data. That data, like all clinical data, is dominated by normal tissue and common, solid, smooth-edged findings. A denoiser trained this way learns what typical anatomy looks like. When it meets a rare lesion, such as a faint ground-glass nodule (a hazy, low-contrast opacity) or a nodule with thin spiculated spokes at its margin, it may "correct" that lesion toward the typical patterns it has seen. In the worst case it turns a malignant-looking nodule into a benign-looking one. We propose to measure this directly, to show that the distortion is caused by imbalance in the denoiser's training data, and to fix it.

The fix rests on a simple observation. For restoration, paired training data for rare lesions can be manufactured with a ground truth that is exact by construction. We insert a synthetic rare-phenotype lesion into a normal-dose scan, which gives the clean target. We then simulate the low-dose acquisition in the projection (sinogram) domain, which gives the noisy input. This sidesteps the usual objection to synthetic augmentation for rare classes, namely that the synthetic examples might carry wrong labels. A denoiser needs no diagnosis label. It only needs to have seen enough rare morphologies to stop treating them as noise.

Core Claim

Existing denoisers optimize fidelity averaged over pixels and over a training set whose lesion phenotypes are long-tailed. Our first claim is that their errors on lesions are therefore concentrated on the rare phenotypes, beyond what lesion size and contrast would predict. Concretely, for each phenotype $k$ (solid, part-solid, ground-glass; smooth, lobulated, spiculated), define the phenotype fidelity gap

$$\Delta_k = E_k^{\text{margin}}(\hat{x}) - E_{\text{solid,smooth}}^{\text{margin}}(\hat{x}),$$

where $\hat{x}$ is the denoised image and $E^{\text{margin}}_k$ is the mean absolute error in Hounsfield units inside a thin band around the boundaries of lesions of phenotype $k$, matched for diameter and contrast. We predict that $\Delta_k$ grows as phenotype $k$ becomes rarer in the denoiser's training set, and grows more for generative denoisers than for regression-style ones.

Our second claim is that rebalancing the denoiser's training distribution by phenotype closes most of this gap without hurting global image quality. The rebalancing uses physics-consistent synthetic lesion insertion. Weighting lesion pixels more heavily in the loss does not close the gap, because reweighting cannot supply morphologies the training set lacks.

Relation to Prior Work

Two papers motivate the claim by showing that globally optimized restoration damages lesions. LIDnet adds a region-proposal network and a perceptual loss on candidate lesion boxes, trained jointly with a detector, and reports better detection after denoising. LeqMod upweights lesion-containing patches in low-count PET denoising and adds uptake-consistency terms. Both treat imbalance as lesion versus background: they make the denoiser look harder at lesions in general. Neither asks which kinds of lesion are lost, and neither changes which morphologies the denoiser has seen. Our proposal targets that phenotype-level imbalance inside the lesion class. We use their designs as our strongest baselines.

Two papers constrain the design by showing that rare morphology carries the diagnosis and is fragile under low-dose noise. Lei et al. show that malignancy classification on low-dose CT depends on local margin cues such as spiculation and lobulation, which global pooling washes out. No Surprises shows that nodule classifiers fail on rare, faint nodule types under screening noise, and repairs this with adversarially searched synthetic nodules. Both work downstream on raw low-dose images. We work on the restoration step, which radiologists and downstream models both consume.

Two papers supply methods we reuse. LeFusion inpaints lesions with controlled shape and texture into real healthy scans. We use it to generate rare-phenotype lesions in normal-dose images, but add the projection-domain noise simulation it lacks. The cone-beam CT diffusion study of Hwang et al. observes that generative translation hallucinates small structures such as minor bronchi. That observation is why we measure hallucination alongside erasure.

One paper is in tension with our design. CT-Mamba enforces normal-dose noise texture globally through a noise-power-spectrum loss. We test whether this global texture prior is itself a channel through which ground-glass texture gets normalized.

Research Program

The work has three parts.

The first part is an audit. We train or reproduce several denoiser families on the 2016 Mayo low-dose challenge data: a convolutional baseline; DenoMamba and CT-Mamba as state-space models; TED-Net as a transformer; an ALDEN-style anatomy-aware adversarial model; and a diffusion model. We evaluate them on LIDC-IDRI chest scans. To make clean/noisy pairs there, we forward-project each scan, add Poisson and electronic noise, and reconstruct. Every annotated nodule is stratified by the radiologist-rated texture and margin morphology that Lei et al. rely on.

Beyond $\Delta_k$, we use a frozen probe. A morphology classifier trained only on normal-dose nodules is applied to each denoised nodule, and we record how often the denoiser flips the probe's verdict: ground-glass to solid, or spiculated to smooth. Flip rate is the clinically meaningful version of the gap.

The second part is a causal test. Correlation between rarity and error could be confounded by size or contrast. So we train the same denoiser on controlled mixtures in which the share of spiculated or ground-glass lesions in the training pairs is set at roughly zero, one, five, and twenty percent. Lesions are inserted into Mayo and LIDC normal-dose backgrounds, and everything else is held fixed. If imbalance in the restoration prior is the cause, flip rate and $\Delta_k$ should fall steadily with training frequency.

The third part is the fix: phenotype-balanced training. We generate rare-phenotype lesions with shape and texture control in normal-dose space and simulate their low-dose versions in the projection domain. We then train denoisers on a mixture balanced by phenotype rather than by lesion presence.

Baselines

The naive reference is the low-dose input itself, together with each denoiser trained on its standard objective with no lesion awareness. Every improvement is measured against this, and it also tells us whether denoising helps or hurts rare-phenotype readability at all.

The strongest restoration competitor is lesion-focused optimization in the style of LIDnet: an ROI perceptual loss on detector-proposed boxes with alternating denoiser–detector training. The closely related second competitor is LeqMod-style lesion-patch oversampling with a lesion-probability consistency term. These are the strongest because they already embody "pay more attention to lesions," so they are the natural answer to the imbalance question.

The reported number for the LIDnet idea comes from its own paper: on the simulated lung carcinoma set L-CT-G, collaborative training raised detection AP-50 for a MAP-NN denoiser from 92.84 to 95.14 (LIDnet). No reported result was found on Lacuna for LeqMod on transmission CT; its evaluation is on PET. Both baselines will be reimplemented on the same backbone, the same simulated dose levels, and the same number of training iterations. They get the same hyperparameter-search budget as our method, and their lesion weighting is tuned on a validation split. Beating them means a statistically significant reduction in rare-phenotype flip rate and $\Delta_k$ at matched global PSNR and SSIM on Mayo2016, and no worse common-phenotype fidelity.

For global quality, we report each backbone's published operating point where the file has one. CT-Mamba reports PSNR 46.2251 and SSIM 0.9805 on Mayo2016 under its own protocol, and ALDEN reports PSNR 33.71 under its protocol. We compare within our own unified protocol rather than across these.

Downstream, the strongest imbalance-aware detector is No Surprises. It reports a competition performance metric of 87.89% on NLST, against 80.62% for its baseline. We run a detector trained this way on raw versus denoised inputs.

The key-ingredient ablation compares phenotype-balanced real-lesion oversampling, phenotype-balanced synthetic insertion with image-domain noise, and phenotype-balanced synthetic insertion with projection-domain noise. All three use the same total number of lesion pairs.

Why not just…?

Why not just reweight lesion pixels in the denoising loss? This is LIDnet's and LeqMod's approach. It raises the cost of error on whatever lesions the model sees, but a spiculated ground-glass nodule that appears a handful of times still contributes almost nothing to what the denoiser learns about such morphology. The causal-frequency experiment settles this directly. If lesion-weighted training at one percent rare-phenotype frequency matches phenotype-balanced training, our claim is wrong, and we would report reweighting as sufficient.

Why not skip denoising and train the downstream model on raw low-dose images with an imbalance-aware loss? Options include Balance Loss, which reports kidney-tumor Dice of 66.8 versus 59.8 for cross-entropy, or No Surprises augmentation. This is a real threat for automated pipelines, and we test it. Still, denoised images are what radiologists read, so a denoiser that erases spiculation harms human reading whatever a robust detector does. We also report whether the denoised-plus-robust-detector pipeline beats the raw-plus-robust-detector pipeline on rare-phenotype sensitivity. If it does not, we will state that the contribution is confined to human-read images.

Why not just oversample the real rare-phenotype nodules already in LIDC-IDRI? This is the simplest fix and may work well, because paired data for real nodules can be simulated just as easily. It is the first arm of our ablation. We expect synthesis to add value mainly for phenotype combinations that are nearly absent in the data, and for avoiding memorization of a few specific nodules. We test this on a held-out set of rare-combination nodules that neither arm trains on. If oversampling matches synthesis there, we will recommend oversampling and drop the generator.

Why not paste lesions and add noise in image space? This would be cheaper than projection-domain simulation. However, image-space noise lacks the streaks and non-stationary mottle of real photon starvation. The image-domain arm of the ablation, evaluated on real low-dose NLST scans rather than simulated ones, shows whether this matters.

Proposed Experiments and Decisive Checks

The audit and causal experiments run on Mayo2016 for training and on LIDC-IDRI with projection-domain dose simulation at roughly 25% and 10% dose for evaluation. Each lesion is matched on diameter and contrast across phenotypes. Results are reported as $\Delta_k$, probe flip rate, and margin-band error, with bootstrap confidence intervals over patients.

Hallucination is measured symmetrically. We run every denoiser on lesion-free regions and count probe-detected lesion-like structures that are absent in the normal-dose target. We also run held-out insertions of small vessels and airways, in light of the bronchus hallucination reported by Hwang et al.. A fix that reduces erasure by increasing hallucination counts as a failure.

Global quality on Mayo2016 must stay within a pre-registered margin of each unmodified backbone. Final validation uses real NLST low-dose scans, where no clean target exists. There we use downstream rare-phenotype sensitivity of a frozen normal-dose-trained classifier and the No Surprises-style detector, plus a small blinded reader comparison of margin conspicuity.

Three outcomes are decisive. Flip rate must fall monotonically with training frequency in the controlled mixtures. Phenotype-balanced training must beat both lesion-weighted baselines on held-out rare-combination nodules at matched PSNR. And the projection-domain arm must beat the image-domain arm on real NLST scans. Failure of the first outcome refutes the premise. Failure of the second or third reduces the method to its simpler alternative, and we would report it that way.