When a medical dataset has eighty glioblastomas for every eight oligodendrogliomas, the classifier sees too few examples of the rare tumor type to learn what distinguishes it. The usual fix is to generate synthetic examples of the rare class. That approach hits two walls. First, a generator trained on a handful of rare cases tends to copy them, so the classifier memorizes the same few patients again. Second, the benefit flattens out: adding more synthetic images stops helping beyond about ten synthetic per real one. This proposal starts from an observation in radiomics, the practice of extracting hand-designed texture statistics from scans. In that work, the signal separating tumor grades and subtypes sits mostly in fine-scale, direction-specific texture. That texture is captured by a wavelet transform, which splits an image into a coarse low-resolution copy plus several "detail" bands that each hold high-frequency edges in one orientation. The overall anatomy, by contrast, is largely shared across classes.
The idea is to give the scarce minority data a narrower job. A generator does not need to invent whole rare-class volumes from scratch. Instead, we take real scans from any patient, majority or minority, keep the wavelet bands that carry little class information, and regenerate only the class-informative bands under the rare-class label. Anatomical diversity then comes from the large majority pool, and the rare class only has to teach texture. We expect this to reduce memorization, to keep paying off past the 10:1 plateau, and to beat both whole-volume synthesis and loss-reweighting on minority-class accuracy. If the measured class information turns out not to concentrate in a subset of bands, the method collapses to ordinary conditional synthesis. We test that condition first.
Core Claim
In imbalanced medical image classification, synthetic minority samples help in proportion to how faithfully they reproduce the minority class's wavelet detail-band statistics without duplicating specific minority patients. A diffusion model that works directly on wavelet coefficients can exploit this. Diffusion models are generators that learn to turn random noise into samples by gradual denoising; this one is trained to resample only the class-informative bands of an existing real volume, conditioned on a target class. Applied to the adult glioma subtype task, we predict it will raise balanced accuracy and minority recall over the strongest loss-based method and over class-conditional whole-volume synthesis. We also predict that its gains will keep rising at synthetic-to-real ratios where prior augmentation saturated.
The new ingredient is the band-selective factorization: what to keep, what to regenerate, and how that choice is measured. Using wavelets as the generator's working space is not new, and neither is class-conditioning a generator.
Where This Sits
The method reuses the representation from two papers. WDM: 3D Wavelet Diffusion Models for High-Resolution Medical Image Synthesis trains diffusion directly on 3D wavelet coefficients instead of on a learned autoencoder's compressed code. This lets it generate full 128³ and 256³ volumes on one 40 GB GPU, but it generates unconditionally and treats all bands alike. cWDM: Conditional Wavelet Diffusion Models for Cross-Modality 3D Medical Image Synthesis adds conditioning by concatenating the wavelet coefficients of the available MRI sequences to the noisy target. It is scored only on reconstruction quality (PSNR 27.31–29.74, SSIM 0.935–0.956 on BraTS 2024) and never on a downstream task. We borrow cWDM's concatenation-style conditioning. The difference is that the condition is a subset of bands from the same volume, and the label being imposed differs from the source patient's label. Neither paper asks which bands carry diagnostic information. The closest bridge to classification is Functional MRI Time Series Generation via Wavelet-Based Image Transform and Spectral Flow Matching for Brain Disorder Identification, which generates fMRI signals in a wavelet-then-cosine space and lifts MDD accuracy from 58.9% to 70.8%. It works on 1D time series with near-balanced cohorts and regenerates the whole signal.
Three papers motivate the band-selective design. Predicting histological grade in pediatric glioma using multiparametric radiomics and conventional MRI features found wavelet features made up 9 of the top 10 predictors of glioma grade. It also found that coarse morphological signs, such as ring enhancement and diffusion restriction, still added information. That second finding warns us that the coarse band may not be class-neutral, which is why the band choice is measured rather than assumed. Augmenting medical image classifiers with synthetic data from latent diffusion models documents the 10:1 saturation plateau. Synthesizing Diverse Lung Nodules Wherever Massively: 3D Multi-Conditional GAN-based CT Image Augmentation for Object Detection reports that going from 1× to 3× synthetic nodules hurt detection. Our hypothesis is that both plateaus partly reflect generators that get fine texture wrong or copy it.
Two papers constrain the design. WaveDIF: Wavelet sub-band based Deepfake Identification in Frequency Domain shows that generators leave detectable fingerprints in specific wavelet bands, including the coarse band and the high-frequency bands. Regenerated bands could therefore carry artifacts a classifier learns to exploit, and we must check this directly. Optimizing Synthetic Data for Enhanced Pancreatic Tumor Segmentation found that boundary noise in synthetic lesions dropped U-Net Dice from 48.48% to 32.49%. Keeping the real coarse band and the real tumor mask fixed is partly a response to that result: lesion geometry stays real.
The comparator on the imbalance side comes from Imbalance-aware loss functions improve medical image classification. That paper combines cross-entropy with differentiable soft-F1 or soft-MCC losses plus oversampling on an 80/12/8 adult glioma cohort. It reports up to 12.7% relative improvement in balanced accuracy. The authors also note that oversampling risks memorizing minority cases, which is the failure our method targets.
Research Program
The first stage measures where class information lives in the wavelet bands. We decompose each multiparametric MRI volume (T1w, T2w, FLAIR) with a two-level 3D Haar transform. Each level yields one coarse band and seven detail bands. We then train a small 3D classifier on real data alone, under band ablations: each band, or each group of bands, is replaced by its class-agnostic population mean. For band $b$ we define its informativeness as
$$I_b = \text{BAcc}(\text{all bands}) - \text{BAcc}(\text{band } b \text{ replaced by its population mean}),$$
where BAcc is balanced accuracy on held-out real data. We average $I_b$ over cross-validation folds. Band selection uses training folds only. The regenerated set $R$ is the bands whose $I_b$ is significantly above zero; the kept set $K$ is the rest.
The method stands or falls here. If $R$ turns out to be every band, there is nothing to factorize. If $R$ includes the coarse band, as the pediatric glioma result hints might happen for morphology, we move to a second decomposition level and keep only the coarsest band. We report the outcome either way.
The second stage builds the generator. We pretrain a wavelet diffusion model in the WDM style on BraTS 2023 and 2024 volumes for anatomical and texture priors. We then fine-tune it on the glioma subtype training split. The model denoises only the bands in $R$. It is conditioned on the real bands in $K$, the tumor mask, and the subtype label. Minority classes are upweighted during fine-tuning. At augmentation time we take any training volume, whether glioblastoma, astrocytoma, or oligodendroglioma, and keep its $K$ bands and mask. We sample new $R$ bands under the minority label, then invert the wavelet transform.
Because the source volumes come from the large majority pool, the number of distinct anatomies available for rare-class synthesis grows by roughly an order of magnitude. Two safeguards apply. A memorization audit computes, for each synthetic sample, the nearest-neighbor distance in $R$-band space to every real minority patient, and compares it to the same distance between pairs of real minority patients. A label-plausibility check runs an independent classifier, trained on real data only, over the synthetic samples. Samples it assigns strongly to the source class are flagged. We report their fraction rather than silently discarding them, since aggressive filtering could hide label noise.
The third stage trains downstream classifiers. We use a 3D convolutional classifier with cross-entropy plus soft-F1 loss, so that our method is layered on the strongest loss-based recipe rather than competing against it. We sweep synthetic-to-real ratios for the minority classes at 1, 3, 10, 30, and 75.
Baselines
The naive reference point is a 3D classifier trained with plain cross-entropy on the imbalanced real data with standard geometric augmentation. It shows how much the imbalance costs.
The strongest existing method is the cross-entropy plus soft-F1 loss with minority oversampling from Imbalance-aware loss functions improve medical image classification. It is the strongest competitor because it was developed on this exact three-class adult glioma imbalance. That paper reports a relative balanced-accuracy improvement of up to 12.7% over its baseline. The file gives no absolute balanced-accuracy value on Lacuna for this dataset, so we will reproduce it ourselves, with the same architecture, splits, and TCGA held-out test, before claiming anything. Beating it means a higher mean balanced accuracy and oligodendroglioma recall across five seeds, with a paired bootstrap interval excluding zero. We also require that majority-class recall does not drop by more than a pre-registered two points.
The strongest generative competitor is the obvious idea at this intersection: a class-conditional wavelet diffusion model that generates whole minority volumes, built from WDM with cWDM-style conditioning on label and mask. It shares our backbone, pretraining, and sampling budget. The only difference is that it regenerates every band from noise instead of keeping $K$ from a real source. We also run a latent-autoencoder diffusion baseline in the spirit of Augmenting medical image classifiers with synthetic data from latent diffusion models, adapted to 3D, because it is the standard synthetic-augmentation recipe.
Every generative arm receives the same pretraining data, the same GPU-hour budget for fine-tuning, the same number of downstream training steps, and the same classifier hyperparameter search of twenty trials. Two ablations isolate the key ingredient. The first regenerates a random band set of the same size as $R$, which tests whether the informativeness measurement matters. The second uses our selective method but restricts source volumes to minority patients only, which tests whether borrowing majority anatomy is what helps.
Why not just…?
Why not just use the imbalance-aware losses and skip generation? This is a serious alternative, and it is also our base recipe. Its own authors concede that oversampling risks memorization, and loss reweighting can only reuse the minority examples that already exist. On natural-image benchmarks, Deep Generative Model for Robust Imbalance Classification reports minority recall of 0.2987 for generative perturbation versus 0.1063 for a Bayesian CNN at imbalance ratio 100. That result is suggestive but comes from CelebA, not MRI. The direct test is our comparison of loss-only against loss plus selective synthesis. If the gap is within seed noise, we will report that generation is not worth its cost here.
Why not just generate whole minority volumes with a conditional wavelet diffusion model? This is the obvious combination of WDM, cWDM, and class conditioning, and it is the method most likely to make ours redundant. Our argument is that with 8% of a modest cohort, a whole-volume generator either copies minority patients or produces texture that blurs toward the majority. The memorization audit and the ratio sweep settle this. If whole-volume synthesis matches our nearest-neighbor distances and keeps improving past 10:1, the band-selective idea adds nothing and we will say so.
Why not just low-pass filter the synthetic images to strip generator artifacts? WaveDIF shows generator fingerprints live partly in high-frequency bands, so filtering them out is a cheap fix. But the radiomics evidence says that is exactly where the minority signal lives, so filtering should erase the reason for synthesizing at all. We include a filtered whole-volume arm. We predict it lowers artifact detectability but also lowers minority recall. If it does not, that is a genuine finding against our premise.
Why not just use mixing augmentations that combine images from different classes? Interpolating whole volumes between classes creates samples with ambiguous texture and labels. Our method also crosses classes, but only in bands measured to be class-neutral. The random-band ablation addresses this: if any band split works as well as the measured one, the gain comes from generic mixing rather than from where class information lives.
Proposed Experiments
The primary benchmark is adult glioma subtype classification with the 80/12/8 split and TCGA held-out testing used by the imbalance-loss paper. BraTS 2023 and 2024 serve only for generator pretraining, and we check that no patients overlap with the test set. Metrics are balanced accuracy, per-class recall and F1, and macro-averaged MCC on the TCGA test set, all over five seeds.
The secondary benchmark tests whether the mechanism transfers beyond MRI and beyond classification. On LIDC-IDRI we treat ground-glass nodules as the rare class and apply selective regeneration to nodule patches inserted into real CT. We compare against the 3D multi-conditional GAN's reported detection score: 0.518 for real data alone and 0.550 with 1× synthetic nodules, measured by competition performance metric with 3D Faster R-CNN. We reproduce the real-only baseline before comparing. Generalization to the 2D dermatology benchmarks is left as an optional check against the reported 76.9% to 84.4% accuracy gain on the darkest skin types.
Four checks decide the outcome. The first is the band-informativeness result: does class information concentrate in a strict subset of bands? If it does not, the project changes course. The second is the ratio sweep. Our method should show a minority-recall curve still rising between 10:1 and 30:1 while whole-volume synthesis flattens; a flat curve for our method falsifies the saturation hypothesis. The third is the memorization audit. Selective samples should sit no closer to real minority patients than real minority patients sit to one another, while whole-volume samples should sit closer. The fourth is an artifact probe. We train a WaveDIF-style detector to separate real from synthetic within the regenerated bands, then retrain the downstream classifier with that detector's signal removed by matching band statistics. If the minority gains vanish, the classifier was exploiting generator fingerprints rather than learning pathology, and we will report the result as negative.
References
- WDM: 3D Wavelet Diffusion Models for High-Resolution Medical Image Synthesis
- cWDM: Conditional Wavelet Diffusion Models for Cross-Modality 3D Medical Image Synthesis
- Functional MRI Time Series Generation via Wavelet-Based Image Transform and Spectral Flow Matching for Brain Disorder Identification
- Predicting histological grade in pediatric glioma using multiparametric radiomics and conventional MRI features
- Augmenting medical image classifiers with synthetic data from latent diffusion models
- Synthesizing Diverse Lung Nodules Wherever Massively: 3D Multi-Conditional GAN-Based CT Image Augmentation for Object Detection
- WaveDIF: Wavelet sub-band based Deepfake Identification in Frequency Domain
- Optimizing Synthetic Data for Enhanced Pancreatic Tumor Segmentation
- Imbalance-aware loss functions improve medical image classification
- Deep Generative Model for Robust Imbalance Classification