Localized Augmentation in Deep Learning

• Localized augmentation is a method that confines transformations to task-relevant regions, enhancing realism by altering only essential parts of the data.
• It improves model performance by preserving unaltered regions while applying targeted corrections, as demonstrated in MRI artifact reduction and lesion inpainting.
• It encompasses diverse approaches such as spatial confinement, neighborhood-conditioned augmentation, and domain-local invariance modeling to address various challenges.

Localized augmentation denotes a class of augmentation procedures in which synthesis, deformation, perturbation, or restoration is restricted to task-relevant subregions, local neighborhoods, or structured local supports rather than applied uniformly to the full sample. In the deep learning-based MRI artifact reduction model DMAR, “localized” is defined in two coupled senses: synthetic motion artifacts are confined to selected subregions of an otherwise clean slice during augmentation, and correction is applied only inside artifact boxes detected at inference time (Zhao et al., 2020). Subsequent work generalized the same principle to anatomy-informed deformation fields in prostate MRI (Kovacs et al., 2023), lesion-level populate/inpaint operators in segmentation (Basaran et al., 2023), local gamma transforms inside ischemic stroke masks (Middleton et al., 2024), local manifold sampling for self-supervised learning (Yang et al., 2022), node-neighborhood feature generation for graph neural networks (Liu et al., 2021), and physics-aware per-access-point perturbations in CSI localization (Serbetci et al., 2022). The term accordingly refers less to a single operator than to a recurring design principle: locality is imposed so that augmentation changes the part of the signal that matters while preserving the remainder.

1. Conceptual forms of locality

A cross-domain synthesis suggests that published localized augmentation methods instantiate locality in at least three operational forms. The first is spatial confinement, where only a subset of pixels or voxels is altered. DMAR pastes artifacted circular ROIs into clean MRI slices and later restricts CAE-based correction to SSD-detected boxes (Zhao et al., 2020). LesionMix pastes or removes lesion volumes while leaving the rest of the 3D image intact (Basaran et al., 2023). Local gamma augmentation modifies DWI intensities only inside the lesion mask, using

$$I_{M,\gamma}(x) = (1 - M(x)) \cdot I(x) + M(x) \cdot I_\gamma(x),$$

so non-lesion voxels are unchanged (Middleton et al., 2024).

The second is neighborhood- or structure-conditioned augmentation, where the transformation is defined by a local manifold, adjacency pattern, or anatomy-derived field. Local Manifold Augmentation constructs $\mathcal{N}_k(x)$ in SSL feature space and samples new views from an instance-conditioned generator anchored to that neighborhood (Yang et al., 2022). LA-GNN learns the conditional distribution $p(x_{\text{neighbor}} \mid x_{\text{central}})$ and injects sampled neighbor features into the first GNN layer (Liu et al., 2021). Anatomy-informed prostate MRI augmentation defines a smooth deformation field

$$F(x) = x + u(x), \qquad u(x) = \nabla(G_\sigma * S_{\text{organ}})(x)\cdot C,$$

so the induced warp decays away from the rectum or bladder boundary (Kovacs et al., 2023).

The third is domain-local invariance modeling, where “local” does not refer to image coordinates but to physically meaningful subsystems. In CSI localization, Phase-DA multiplies all subcarriers and antennas of access point $k$ by a common factor $e^{j\phi_k}$, and Amp-DA scales them by a common $\alpha_k$; locality is per access point rather than per pixel (Serbetci et al., 2022). LocaGen similarly treats unseen coordinates as local targets in a spatial field and weights training losses by distances between seen and unseen locations (Abdelmotlb et al., 22 Nov 2025).

This broad usage is important because it distinguishes localized augmentation from generic geometric or intensity transforms. The unifying property is selective support, not any single image-space implementation.

2. Localized augmentation in DMAR and motion-corrupted brain MRI

The most explicit early formulation appears in “Localized Motion Artifact Reduction on Brain MRI Using Deep Learning with Effective Data Augmentation Techniques” (Zhao et al., 2020). DMAR uses a two-stage architecture: SSD detects localized artifact boxes with confidence scores, and a convolutional autoencoder reconstructs artifact-free content only inside those detected boxes. The selective second stage is intended to leave unaffected areas untouched and thereby prevent overcorrection.

Its training pipeline is built around localized synthesis. Starting from OASIS-1 T1-weighted MRI scans, 375 scans were used for training. Per scan, 50 slices were sampled from each plane—sagittal, axial, and coronal—yielding 150 slices per scan and a total of 56,525 base slices for augmentation (Zhao et al., 2020). Three augmentation mechanisms were then applied. First, motion-free images were diversified by simulating inter-subject morphological variability: each slice was locally warped inside 3–8 non-overlapping circles using

$$IMG_{\text{new}}(P) = IMG_{\text{old}}(C + u^{1+\epsilon}(P-C)),$$

with $u=\text{distance}(P,C)/R$ and $\epsilon=0.2$ (Zhao et al., 2020). Second, ringing-like artifacts were generated by modifying annular sectors in k-space, inverting the Fourier transform, and then localizing the corruption by copying random circular ROIs from the globally artifacted image back into the clean slice, followed by intensity histogram matching. Third, “rippling” artifacts were produced by elliptic sine-based intensity modulation and then similarly localized by ROI paste-back (Zhao et al., 2020).

Localization is central both to supervision and to inference. SSD labels are generated automatically as the circumscribing squares of the circular ROIs used during augmentation, so the detector is trained directly on local artifact extents. Ringing and rippling artifacts are combined in a 2:1 proportion to reflect observed prevalence in ABIDE exemplars, and the full pipeline yields 225,000 paired samples for SSD and CAE training (Zhao et al., 2020). Quantitatively, depending on degradation level, DMAR achieves a 27.8%-48.1% reduction in RMSE and a 2.88--5.79 dB gain in PSNR on a 5000-sample synthetic test set. On 55 real-world motion-affected slices from 18 ABIDE subjects, the model reduced the variance of image voxel intensity within artifact-affected brain regions with $\mathcal{N}_k(x)$0 (Zhao et al., 2020).

DMAR established two ideas that recur in later literature: local corruption is often easier to synthesize realistically than whole-image corruption, and local restoration is often safer than global post-processing.

3. Medical-imaging variants: anatomy, lesions, and pathology-specific intensity models

Medical imaging has become the main domain in which localized augmentation is operationalized as an anatomical prior rather than a generic perturbation. In prostate MRI, anatomy-informed augmentation uses rectum and bladder masks predicted by an nnU-Net model and quality-checked by radiologists, with crops extending $\mathcal{N}_k(x)$1 mm axial to the prostate and $\mathcal{N}_k(x)$2 mm in-plane from rectum or bladder (Kovacs et al., 2023). The deformation field is applied consistently to all registered sequences and to lesion masks by nearest-neighbor resampling, so label correspondence is preserved. Validation selected $\mathcal{N}_k(x)$3 and $\mathcal{N}_k(x)$4, with $\mathcal{N}_k(x)$5 and augmentation probability $\mathcal{N}_k(x)$6 (Kovacs et al., 2023). On 774 biopsy-confirmed examinations, rectum + bladder deformation improved F1 by +5.11% over basic augmentation at the selected working point and detected 43/76 lesions versus 39/76, with $\mathcal{N}_k(x)$7 (Kovacs et al., 2023). The same study also showed that random elastic deformation could improve lesion-level behavior marginally while degrading patient-level pAUROC, underscoring that anatomical plausibility, not just variability, governs utility.

Lesion-level methods represent another branch. LesionMix performs lesion populating and lesion inpainting in 3D, guided by a lesion likelihood map and a target lesion volume distribution $\mathcal{N}_k(x)$8 (Basaran et al., 2023). Its populate operator is

$\mathcal{N}_k(x)$9

with soft boundary blending in the image and hard lesion labels in the mask (Basaran et al., 2023). On WMH2017 at full training size, LesionMix reached Dice 80.95 versus 79.13 for TDA, 80.14 for CutMix, and 79.74 for CarveMix; on LiTS it achieved 63.51 versus 61.93 for TDA and 58.39 for CutMix (Basaran et al., 2023). CarveMix, by contrast, derives a lesion-aware ROI from the signed distance function of the source mask and pastes that ROI into a second volume:

$p(x_{\text{neighbor}} \mid x_{\text{central}})$0

Its gains are especially large in low-data regimes; on the ATLAS dataset at 12.5% training size, CarveMix achieved 54.77 Dice versus 41.86 for TDA and 24.57 for CutMix (Zhang et al., 2021).

A third subgroup modifies lesion appearance in situ rather than moving lesions across subjects. Local gamma augmentation for ischemic stroke uses a lesion-specific min–max normalization and samples

$p(x_{\text{neighbor}} \mid x_{\text{central}})$1

then applies the gamma transform only inside the lesion mask on DWI (Middleton et al., 2024). The method targets underrepresentation of DWI-positive/FLAIR-negative hyperacute presentations. Relative to a baseline U-Net already trained with global gamma and other standard transforms, local gamma increased image-level sensitivity from 0.781 to 0.844 on MVS, from 0.757 to 0.814 on RH, and from 0.566 to 0.736 on WUS, with slight specificity reductions on MVS and RH (Middleton et al., 2024). For chest X-ray nodule classification, lung inpainting-based local feature augmentation removes nodules from 64×64 patches, extracts the nodule by subtraction, and reinserts it at arbitrary locations inside lung masks with local rotation and flipping (Guendel et al., 2020). At full training size, AUC improved from 0.792 ± 0.010 to 0.805 ± 0.004; at 5% training size, it improved from 0.649 ± 0.009 to 0.669 ± 0.013 (Guendel et al., 2020).

These variants share a precise premise: the pathological region is sparse and semantically privileged, so augmentation should alter that region directly rather than dilute supervision through whole-image transforms.

4. Locality in manifolds, features, and graph neighborhoods

Localized augmentation is not limited to the input domain. In self-supervised vision, Local Manifold Augmentation models the empirical local manifold around an image using $p(x_{\text{neighbor}} \mid x_{\text{central}})$2-nearest neighbors in SSL feature space, trains an instance-conditioned generator, and defines

$p(x_{\text{neighbor}} \mid x_{\text{central}})$3

with LMA applied before otherwise standard handcrafted augmentations (Yang et al., 2022). The stated motivation is that cropping, flipping, color jitter, and blur do not cover complex intra-class variation such as pose, viewpoint, lighting condition, and background. With LMA, SimSiam on ImageNet100 improved from 78.32 to 82.94 and MoCov2 from 69.80 to 80.80; on ImageNet, SimSiam improved from 65.62 to 67.82 and MoCov2 from 62.48 to 63.97 (Yang et al., 2022).

For graph neural networks, locality is defined by the 1-hop neighborhood rather than geometric support. LA-GNN trains a CVAE on observed neighbor pairs and learns the conditional distribution of neighbor features given the central node feature. At each training iteration it samples $p(x_{\text{neighbor}} \mid x_{\text{central}})$4 augmented feature matrices and injects them into the backbone’s first layer by concatenation or averaging (Liu et al., 2021). On Cora, Citeseer, and Pubmed, plugging local augmentation into GCN improves by an average of 3.4% in test accuracy, while GAT improves by an average of 1.6% (Liu et al., 2021). The paper further reports that gains are larger for low-degree nodes, consistent with the claim that local augmentation primarily benefits sample-starving neighborhoods.

Local Magnification adds a different feature-space interpretation. LOMA magnifies a random local area of the image via an inverse radial warp inside a rhombus or ellipse and extends the same idea to intermediate feature maps, yielding LOMA_IF and FO (He et al., 2022). On CIFAR-100 with WideResNet-28-10, baseline accuracy 81.32 ± 0.20 increased to 82.28 ± 0.23 with image-space LOMA, 82.65 ± 0.30 with LOMA_IF, and 82.89 ± 0.23 with LOMA_IF&FO (He et al., 2022). A related but broader usage appears in Local Model Feature Transformations, where neighborhoods are used to fit local Gaussian processes, quadrics, linear classifiers, or word embeddings, and the resulting parameters, uncertainties, residuals, or geometric descriptors are concatenated to raw inputs as augmented representations rather than synthetic samples (Brown, 2020).

A common misconception is that localized augmentation is necessarily pixel-level. These examples show that locality can be imposed on a manifold chart, a node neighborhood, or an intermediate feature tensor without any explicit region-of-interest mask in the input image.

5. Spatially localized augmentation in sensing, localization, and adversarial optimization

In signal and localization problems, localized augmentation often encodes spatial or hardware-local invariances rather than image anatomy. For CSI-based indoor localization, Phase-DA samples $p(x_{\text{neighbor}} \mid x_{\text{central}})$5 independently for each AP and transforms CSI as

$p(x_{\text{neighbor}} \mid x_{\text{central}})$6

while Amp-DA samples per-AP dB offsets $p(x_{\text{neighbor}} \mid x_{\text{central}})$7, sets $p(x_{\text{neighbor}} \mid x_{\text{central}})$8, and applies

$p(x_{\text{neighbor}} \mid x_{\text{central}})$9

The critical restriction is that the phase or gain is constant across all subcarriers and antennas of an AP, preserving AoA/ToF structure (Serbetci et al., 2022). On the WILD dataset, Phase-DA reduced LOS small-data MSE from 0.994347 to 0.307535 at ×6 augmentation and NLOS small-data MSE from 5.204818 to 1.349240 at ×5 augmentation. The same paper reports that a 4k set augmented to match a 40k set can reach comparable performance, yielding 0.823344 m versus 0.824396 m in NLOS (Serbetci et al., 2022).

LocaGen treats locality as spatial support in a coordinate field. It divides target locations into seen and unseen subsets with a density-based initializer, augments seen fingerprints with domain heuristics, and trains a conditional diffusion model with a spatially weighted loss

$$F(x) = x + u(x), \qquad u(x) = \nabla(G_\sigma * S_{\text{organ}})(x)\cdot C,$$0

Its purpose is to synthesize fingerprints at completely unseen coordinates while respecting neighboring seen measurements (Abdelmotlb et al., 22 Nov 2025). On UJIIndoorLoc, LocaGen maintains the same localization accuracy even with 30% of the locations unseen and achieves up to 28% improvement in accuracy over state-of-the-art augmentation methods; the density-based selection itself outperforms grid-center and random seen selection by >31% in mean localization error (Abdelmotlb et al., 22 Nov 2025).

Localized augmentation can also be adversarial rather than robustness-oriented. In ship detection on HRSC2016, augmentation of the whole remote-sensing image was found to introduce background-induced false positives and false negatives unrelated to the adversarial patch, so the proposed method applies augmentations only inside target ship regions:

$$F(x) = x + u(x), \qquad u(x) = \nabla(G_\sigma * S_{\text{organ}})(x)\cdot C,$$1

On YOLOv5-M, global augmentation changed recall from 59.3% to 48.1% and ASR from 42.0% to 49.3%, whereas localized augmentation further reduced recall to 42.7% and increased ASR to 54.1% (Liu et al., 29 Aug 2025). Here locality is used to focus optimization on the patch-target interaction and suppress interference from non-target background.

6. Advantages, failure modes, and recurrent design constraints

The literature consistently associates localized augmentation with three advantages. First, it can increase realism by restricting synthesis to regions where the transformation is physically or anatomically meaningful. This is explicit in prostate MRI, where rectum-informed deformation outperforms generic elastic warps at clinically relevant operating points (Kovacs et al., 2023), and in CSI localization, where per-AP transforms outperform naïve i.i.d. complex Gaussian noise injection (Serbetci et al., 2022). Second, it can preserve labels more faithfully because images and labels are warped or composited coherently; this is a central claim in anatomy-informed deformation, LesionMix, CarveMix, and LA-GNN (Kovacs et al., 2023, Basaran et al., 2023, Zhang et al., 2021, Liu et al., 2021). Third, it is often especially valuable in sparse-target or low-data regimes, as shown by the larger relative gains of CarveMix, LesionMix, local feature augmentation for nodules, and DDPM-based landmark augmentation under reduced supervision (Zhang et al., 2021, Basaran et al., 2023, Guendel et al., 2020, Hadzic et al., 2024).

These gains are not unconditional. Several papers document failure modes. Poor rectum or bladder masks can misguide the deformation field in anatomy-informed DA, and large $$F(x) = x + u(x), \qquad u(x) = \nabla(G_\sigma * S_{\text{organ}})(x)\cdot C,$$2 or small $$F(x) = x + u(x), \qquad u(x) = \nabla(G_\sigma * S_{\text{organ}})(x)\cdot C,$$3 can create sharp gradients or fold-over risk (Kovacs et al., 2023). Amp-DA can degrade performance in large-data regimes when $$F(x) = x + u(x), \qquad u(x) = \nabla(G_\sigma * S_{\text{organ}})(x)\cdot C,$$4 is too large (Serbetci et al., 2022). In ALL with diffusion-generated synthetic image/heatmap pairs, acceptance after MRF and SSM filtering was 40% ± 3% on the FullDataset setting but only 26% ± 5% on ReducedDataset, while VAE samples achieved 3% ± 3% and 0% acceptance, respectively (Hadzic et al., 2024). Local gamma improved sensitivity but slightly reduced specificity on two external stroke datasets (Middleton et al., 2024). Global augmentation can also remain necessary: LOMA performs best when combined with standard crop and flip rather than used as a full replacement (He et al., 2022).

A second misconception is that locality automatically implies easier implementation. In practice, many methods depend on auxiliary structures: lesion masks, organ masks, lung segmentations, heatmaps, density estimates, k-NN neighborhoods, or patch placement rules. They also impose modality-specific interpolation constraints—bilinear or trilinear for images and features, nearest-neighbor for masks, consistent multi-sequence warping after registration, or histogram matching during ROI paste-back (Zhao et al., 2020, Kovacs et al., 2023, Basaran et al., 2023, Hadzic et al., 2024). This suggests that localized augmentation is best understood as a structured augmentation regime whose effectiveness depends on the accuracy of the local support definition.

Taken together, the literature portrays localized augmentation as a technically heterogeneous but conceptually coherent family. Whether implemented as local ROI compositing, organ-driven deformation, lesion-normalized intensity transforms, manifold-conditioned generation, neighborhood feature synthesis, or hardware-local signal perturbation, its central thesis is stable: augmentation is most useful when it perturbs the degrees of freedom that vary in the real task while preserving the rest of the sample.