MI-ND: Adaptive Medical Image Denoising

• MI-ND is an adaptive denoising paradigm that adjusts restoration using image statistics, noise estimators, and structural priors.
• It integrates both traditional methods like contourlet-domain thresholding and deep architectures including transformer-based designs to preserve critical anatomical details.
• Empirical results demonstrate its effectiveness in enhancing metrics like PSNR and SSIM while maintaining edge and feature preservation for accurate diagnoses.

Searching arXiv for the cited MI-ND and related medical image denoising papers. Medical Image Adaptive Denoising Model (MI-ND) denotes a family of medical-image denoising methods in which the restoration process is adjusted by image statistics, estimated noise, structural priors, modality information, or downstream diagnostic constraints rather than by a single fixed denoising rule. In the literature considered here, the designation appears explicitly for an explainable denoising framework that preserves gradient-based feature-importance maps (Dong et al., 2023) and for the noise-adaptive framework "MIND: A Noise-Adaptive Denoising Framework for Medical Images Integrating Multi-Scale Transformer" (Tang et al., 11 Aug 2025). Closely related methods implement the same adaptive principle through contourlet-domain thresholding (Satheesh et al., 2011), convolutional denoising autoencoders trained on small heterogeneous datasets (Gondara, 2016), unsupervised dictionary-plus-residual learning for MRI and CT (Rai et al., 2021), edge-enhanced transformer denoising for low-dose CT (Luthra et al., 2021), iterative denoising and motion-artifact correction for 3D brain MRI (Zhang et al., 2024), two-stage multimodal denoising with self-guided noise attention (Sharif et al., 10 Mar 2025), multi-scale repainting for MRI (Pal et al., 11 Mar 2025), and distributed privacy-aware denoising pipelines (Adebayo et al., 8 Nov 2025).

1. Definition and conceptual scope

MI-ND is best understood as an adaptive denoising paradigm for medical imaging rather than as a single invariant architecture. Across the cited literature, the adaptive element takes several forms: threshold selection from noisy-image statistics, explicit noise-level estimation, residual prediction conditioned on corruption strength, structure-aware attention, iterative stopping criteria, multimodal fusion, or feature-preserving supervision. The shared objective is to suppress noise while preserving diagnostically important structures such as edges, contours, fine anatomical details, low-contrast regions, vessels, lesion boundaries, gray–white matter transitions, and layered tissue structure (Satheesh et al., 2011, Dong et al., 2023, Tang et al., 11 Aug 2025).

A recurring distinction in this literature is between explicit and implicit adaptivity. In the contourlet-based MRI method, the threshold is adaptive because it is computed from noisy-image statistics rather than fixed a priori (Satheesh et al., 2011). In the convolutional denoising autoencoder, by contrast, the method is not adaptive in the sense of a dynamic noise-estimation or modality-switching mechanism; its adaptivity comes from learning data-driven mappings from corrupted to clean images and from combining heterogeneous medical datasets (Gondara, 2016). Later systems move toward explicit conditioning: JDAC estimates noise from gradient-map statistics and conditions a 3D U-Net denoiser on the estimated noise variance (Zhang et al., 2024), while MIND introduces a Noise Level Estimator and a Noise Adaptive Attention Block to realize channel-spatial attention regulation driven by noise perception (Tang et al., 11 Aug 2025).

A useful synthesis is that MI-ND methods differ mainly in where adaptivity is inserted. Some adjust the transform-domain shrinkage rule, some modulate internal features, some alter the iteration depth, and some change the loss landscape so that denoising is constrained by explainability or downstream diagnostic performance. This suggests that MI-ND is less a single model family than a design doctrine centered on noise awareness and structure preservation.

2. Early transform-domain formulations

The earliest formulation in this set is the contourlet-based medical image denoising algorithm for MRI proposed in "Medical Image Denoising using Adaptive Threshold Based on Contourlet Transform" (Satheesh et al., 2011). It targets brain and spine MR images corrupted by additive white Gaussian noise with zero mean. The method uses the contourlet transform because it captures multiscale information, multidirectional features, and contours and edges more effectively than separable wavelets. Its transform is built from a Laplacian Pyramid, which provides multiscale decomposition into low-pass and band-pass outputs, and a Directional Filter Bank, which decomposes the band-pass subbands into multiple directional subbands.

The adaptive step is the thresholding rule in the contourlet domain. Noise variance is estimated using the median absolute deviation from contourlet coefficients,

$$\sigma_n = \frac{\mathrm{median}(C_{i,j})}{0.6745},$$

and the threshold is given in the paper as

$T = 2\sqrt{N}\,\sigma_g^2,$

where $N$ is the total number of pixels in the image and $\sigma_g$ is the standard deviation of the noisy image (Satheesh et al., 2011). Coefficients greater than the threshold are kept unchanged, while coefficients less than the threshold are suppressed; inverse contourlet transform then reconstructs the denoised image. The paper evaluates the method on spine MR images and brain MR images with AWGN at different variances, including examples with variance $=25$ and variance $=30$, and reports higher PSNR than wavelet hard thresholding, soft thresholding, and Wiener filtering.

This early MI-ND-style formulation established several themes that persist later. First, adaptation can be image-statistical rather than learned. Second, medical denoising is treated as structure preservation rather than only variance reduction. Third, contour- and edge-oriented representations were already recognized as important for anatomical boundaries. The paper also exposes a limitation that remains relevant in later work: the threshold derivation is somewhat unclear in notation, and the thresholding rule is relatively simple and may not be optimal for all coefficient distributions (Satheesh et al., 2011).

3. Deep architectures and adaptive mechanisms

Deep MI-ND variants replace fixed transform-domain shrinkage with learned mappings, but they do not converge on a single architecture.

A simple baseline is the convolutional denoising autoencoder for mammograms and dental X-rays (Gondara, 2016). It uses a plain convolutional encoder-decoder stack without explicit skip connections, residual blocks, or U-Net-style lateral connections. Its central result is pragmatic: even with small sample size, denoising autoencoders constructed using convolutional layers can be used for efficient denoising of medical images, and heterogeneous images can be combined to boost sample size. This work is important historically because it argues that high-capacity denoisers are not the only viable route in medical imaging.

The unsupervised framework for MRI and CT combines patch-based dictionary learning and residual learning in an iterative loop (Rai et al., 2021). Overlapping 2D or 3D patches are processed by a K-SVD dictionary branch that indirectly models noise through sparse representation and by a deep residue network that directly estimates the residual. The two residual estimates are averaged and fed back to update the dictionary. This yields a denoiser that is unsupervised, works on both 2D slices and 3D volumes, and explicitly targets modality-dependent noise statistics such as Rician noise in MRI and Poisson noise in CT.

Transformer-based MI-ND variants emphasize structure modeling. Eformer builds an encoder-decoder network using LeWin transformer blocks, non-overlapping window-based self-attention, and learnable Sobel-Feldman operators for edge enhancement (Luthra et al., 2021). The network repeatedly concatenates edge-enhanced maps $S(I)$ with intermediate features, and the residual-learning variant predicts the noise residual rather than the clean image directly. On the AAPM-Mayo Clinic Low-Dose CT Grand Challenge Dataset, the residual variant reports $43.487$ PSNR, $0.0067$ RMSE, and $0.9861$ SSIM (Luthra et al., 2021).

The explainable framework explicitly named MI-ND introduces a pretrained U-Net restoration network, Guided Backpropagation, and a feature-preserving loss $T = 2\sqrt{N}\,\sigma_g^2,$0 that penalizes discrepancies between the XAI feature map of the clean image and that of the denoised image (Dong et al., 2023). Here, adaptation is neither purely architectural nor purely statistical; it is imposed by a supervisory signal that biases the denoiser toward preserving the same explanation structure before and after denoising. The core claim is that ordinary MSE or residual learning handles pixels dominated by noise or by feature content alone, but struggles in pixels where noise and diagnostic structure overlap.

JDAC extends adaptivity to 3D brain MRI with simultaneous noise and motion artifacts (Zhang et al., 2024). Its denoising component is a conditional 3D U-Net-like network with five encoder blocks and five decoder blocks, conditional normalization driven by the estimated noise standard deviation $T = 2\sqrt{N}\,\sigma_g^2,$1, linear interpolation instead of deconvolution, and pruned skip connections that keep only the two highest-resolution skip connections. It is embedded in an alternating iterative-learning framework with a separate anti-artifact U-Net. This architecture reflects a stronger notion of adaptivity: denoising strength is conditioned on estimated noise, and the number of iterations is controlled by an early stopping criterion tied to the estimated noise level.

The two-stage multimodal framework with self-guided noise attention formalizes another branch of MI-ND (Sharif et al., 10 Mar 2025). Stage I is a Noise Estimation Network built from Residual-in-Residual Dense Blocks that learns $T = 2\sqrt{N}\,\sigma_g^2,$2. Stage II is a Reconstruction Network that learns $T = 2\sqrt{N}\,\sigma_g^2,$3 and uses a self-guided noise attention block to emphasize likely corrupt regions in a coarse-to-refine manner. The design is explicitly multimodal, spanning MRI, X-ray, CT, skin lesion images, and microscopy / protein atlas images, and explicitly multi-pattern, spanning Gaussian and speckle noise.

MIND consolidates several of these threads into an end-to-end noise-adaptive architecture (Tang et al., 11 Aug 2025). It combines a multi-scale residual pyramid encoder-decoder, a Transformer cascade module, cross-modal feature fusion across the raw noisy image, a preliminary denoised image, and a gradient map, a Noise Level Estimator, and a Noise Adaptive Attention Block. The NLE produces a noise map $T = 2\sqrt{N}\,\sigma_g^2,$4 and maps it into modulation parameters $T = 2\sqrt{N}\,\sigma_g^2,$5 and $T = 2\sqrt{N}\,\sigma_g^2,$6, while the NAAB applies the affine modulation

$T = 2\sqrt{N}\,\sigma_g^2,$7

before channel and spatial attention. The result is a denoiser whose feature recalibration is explicitly driven by estimated local noise.

4. Noise models, training regimes, and objective functions

Noise modeling in MI-ND research ranges from simple AWGN assumptions to modality-specific and mixed-noise formulations. The contourlet MRI method assumes AWGN with zero mean (Satheesh et al., 2011). The convolutional autoencoder tests Gaussian and Poisson corruption, including settings such as $T = 2\sqrt{N}\,\sigma_g^2,$8, $T = 2\sqrt{N}\,\sigma_g^2,$9, $N$0, $N$1, and Poisson noise with $N$2 and $N$3 (Gondara, 2016). The unsupervised dictionary-plus-residual framework models MRI with a Rician conditional PDF and CT with a Poisson PMF (Rai et al., 2021). The explainable MI-ND study synthesizes 13 different noise/artifact types, including Gaussian noise, Poisson noise, speckle noise, non-central chi-square noise, salt-and-pepper noise, structured noise, thermal noise, magnetic field inhomogeneity noise, chemical shift artifact, motion artifact, wrap-around artifact, and susceptibility artifact (Dong et al., 2023). The multimodal two-stage denoiser explicitly trains on additive Gaussian noise and multiplicative speckle noise (Sharif et al., 10 Mar 2025). MIND uses Gaussian noise, Poisson noise, speckle noise, and motion blur (Tang et al., 11 Aug 2025). The distributed chest X-ray system uses additive Gaussian noise as a lightweight obfuscation mechanism with fixed mean $N$4 and standard deviations $N$5, $N$6, and $N$7 (Adebayo et al., 8 Nov 2025).

Training regimes are equally diverse. Some are fully supervised with synthetic corruption, as in JDAC, where Gaussian noise with standard deviation sampled uniformly from $N$8 is added to 9,544 T1-weighted MRIs from ADNI and the denoiser is trained with an $N$9 loss on the predicted noise image (Zhang et al., 2024). Some are explicitly unsupervised, as in the dictionary-plus-residual framework, which learns noise characteristics directly from noisy data without clean targets (Rai et al., 2021). Some are unpaired-to-paired, as in ADANI, where a generator synthesizes noisy images from unpaired clean and noisy images so that a U-Net denoiser can be trained in a fully supervised manner on generated pairs (Lin et al., 2020). Some emphasize privacy-aware obfuscation and distributed throughput rather than only denoising accuracy (Adebayo et al., 8 Nov 2025).

Objective functions reflect different views of what should be preserved. The contourlet method evaluates with

$\sigma_g$0

where

$\sigma_g$1

(Satheesh et al., 2011). Eformer uses a combination of MSE and multi-scale perceptual loss based on a frozen ResNet-50 feature extractor (Luthra et al., 2021). The explainable MI-ND system adds the Feature Preserving Loss

$\sigma_g$2

to a residual loss, explicitly regularizing the denoiser to preserve the clean image’s feature-importance map (Dong et al., 2023). The multimodal two-stage method uses a mean-square-error-type residual loss for Stage I and an $\sigma_g$3 reconstruction loss for Stage II (Sharif et al., 10 Mar 2025). MIND uses a weighted composite loss combining MSE, SSIM, edge, perceptual, and adversarial terms, with noise-adaptive weights

$\sigma_g$4

(Tang et al., 11 Aug 2025).

A central methodological implication is that adaptivity is often implemented as a conditioning variable in the loss or normalization path rather than only as an architectural branch. This is especially explicit in JDAC, MIND, and the explainable MI-ND framework.

5. Empirical performance and application domains

The empirical scope of MI-ND research is broad, covering MRI, CT, X-ray, PET/CT, mammography, dental radiography, skin lesions, and microscopy (Gondara, 2016, Rai et al., 2021, Dong et al., 2023, Sharif et al., 10 Mar 2025, Tang et al., 11 Aug 2025). The reported metrics are predominantly PSNR, SSIM, RMSE, LPIPS, MSE, $\sigma_g$5, VIFP, F1 score, and ROC-AUC, depending on whether the study emphasizes restoration quality, perceptual fidelity, or downstream diagnostic effect.

Paper	Domain	Reported outcome
(Satheesh et al., 2011)	Brain and spine MRI	Higher PSNR than wavelet-based denoising algorithms
(Luthra et al., 2021)	Low-dose CT	$\sigma_g$6 PSNR, $\sigma_g$7 RMSE, $\sigma_g$8 SSIM
(Dong et al., 2023)	CT and X-ray	Best PSNR and SSIM across most of 13 noise/artifact settings
(Zhang et al., 2024)	3D brain MRI	$\sigma_g$9 dB PSNR and $=25$0 RMSE on MR-ART denoising
(Sharif et al., 10 Mar 2025)	Multimodal medical images	Overall gain of $=25$1 PSNR, $=25$2 SSIM, $=25$3 $=25$4, $=25$5 VIFP, and $=25$6 MSE
(Tang et al., 11 Aug 2025)	Multimodal public datasets	$=25$7 dB PSNR, $=25$8 SSIM, $=25$9 LPIPS, $=30$0 RMSE, $=30$1 F1, $=30$2 ROC-AUC

Several results are especially influential. In the explainable MI-ND study, MI-ND achieves best PSNR and SSIM across all three datasets and most of the 13 noise/artifact settings, with examples such as LIDC chemical shift at PSNR $=30$3, SSIM $=30$4; LiTS magnetic field at PSNR $=30$5, SSIM $=30$6; and RSNA susceptibility artifact at PSNR $=30$7, SSIM $=30$8 (Dong et al., 2023). In the multimodal two-stage system, the full model has $=30$9 trainable parameters, $S(I)$0 G MACs, and about $S(I)$1 ms inference time for a $S(I)$2 image, while outperforming CAE, DAE, ResCNN, DnCNN, MMD, and DRAN-style baselines (Sharif et al., 10 Mar 2025). In MIND, the reported table gives PSNR $S(I)$3 dB, SSIM $S(I)$4, LPIPS $S(I)$5, RMSE $S(I)$6, F1 $S(I)$7, and ROC-AUC $S(I)$8, all best among the listed methods BM3D, NLM, Wiener, DnCNN, FFDNet, SwinIR, and DDPM (Tang et al., 11 Aug 2025).

The distributed denoising study adds a deployment-oriented dimension. On the NIH ChestX-ray14 subset of 15,000 images resized from $S(I)$9 to $43.487$0, U-Net and U-Net++ are evaluated under single-GPU, DataParallel, and DDP + AMP training (Adebayo et al., 8 Nov 2025). At $43.487$1 noise in the single-GPU setting, U-Net reports PSNR $43.487$2, SSIM $43.487$3, LPIPS $43.487$4, while U-Net++ reports $43.487$5, $43.487$6, and $43.487$7. The optimized DDP+AMP pipeline is reported as the fastest, reducing training time by over $43.487$8 versus single-GPU and outperforming DataParallel by over $43.487$9, with only minor accuracy drop in some cases (Adebayo et al., 8 Nov 2025). This indicates that empirical evaluation in MI-ND has expanded from restoration accuracy alone to throughput, scalability, and privacy-aware training scenarios.

6. Misconceptions, limitations, and open directions

One common misconception is that “adaptive” necessarily means a deep noise-estimation network. The literature does not support that narrow reading. The contourlet method is adaptive because the noise level is estimated from the image itself using MAD and the threshold depends on image size and noisy-image statistics (Satheesh et al., 2011). The autoencoder paper is adaptive only in a weaker, learned sense and explicitly lacks modality-specific branching or adaptive gating (Gondara, 2016). The repainting method is adaptive across scales and noise levels but is not a neural network denoiser; it is more accurately described as a classical, optimization-based adaptive denoising and reconstruction pipeline (Pal et al., 11 Mar 2025). This suggests that MI-ND should be defined functionally, not by architecture class.

A second misconception is that better denoising metrics automatically imply better diagnostic preservation. Several papers argue against this simplification. The explainable MI-ND framework was motivated by the claim that classical denoising algorithms often ignore explanatory and critical medical features preservation, which may lead to misdiagnosis and legal liabilities (Dong et al., 2023). MIND therefore measures not only PSNR, SSIM, and LPIPS but also F1 score and ROC-AUC (Tang et al., 11 Aug 2025). The distributed chest X-ray study shows that U-Net++ can have stronger structural fidelity in PSNR and SSIM while showing less favorable LPIPS than U-Net under low and moderate noise levels (Adebayo et al., 8 Nov 2025). This indicates that “best denoiser” depends on the target criterion.

The literature also presents clear technical caveats. The contourlet paper’s threshold notation is somewhat unclear, and its evaluation is limited to a small set of medical images and Gaussian noise (Satheesh et al., 2011). The autoencoder uses $0.0067$0 resized images and a simple architecture without skip connections or residual learning, which limits direct clinical applicability (Gondara, 2016). The unsupervised dictionary-plus-residual method requires careful patch design and empirically chosen $0.0067$1 and $0.0067$2 (Rai et al., 2021). Eformer is evaluated on a single benchmark and resized $0.0067$3 inputs (Luthra et al., 2021). JDAC is trained only on Gaussian noise and generalizes reasonably to Gaussian-like corruptions such as Rician and speckle, but not to salt-and-pepper noise (Zhang et al., 2024). The two-stage multimodal method is computationally heavier than simpler baselines, and U-Net++ in the distributed study may need finer hyperparameter tuning under mixed-precision distributed settings (Sharif et al., 10 Mar 2025, Adebayo et al., 8 Nov 2025).

Future directions are already visible in the surveyed work. The distributed denoising paper proposes attention mechanisms, perceptual losses such as VGG or adversarial losses, human-in-the-loop evaluation with radiologists, and federated learning (Adebayo et al., 8 Nov 2025). JDAC points toward tighter coupling of denoising and artifact correction in 3D MRI (Zhang et al., 2024). MIND suggests that noise-aware attention and cross-modal fusion can be linked directly to downstream diagnosis (Tang et al., 11 Aug 2025). A plausible implication is that the next generation of MI-ND systems will be evaluated not only by restoration fidelity but also by robustness to non-uniform noise, preservation of clinically meaningful explanations, and compatibility with privacy-preserving or distributed clinical workflows.