ak-MPPCA Denoising — EMBC 2026

Conference Paper · Invention Disclosure · 2026 · Toronto, Canada · GE HealthCare Invention Disclosure

Structure-aware Adaptive Kernel MPPCA Denoising for Diffusion MRI

Ananya Singhal · Dattesh Dayanand Shanbhag · Sudhanya Chatterjee
GE HealthCare, Advanced Technology Group, Bengaluru, India

EMBC 2026 · Engineering in Medicine and Biology Society · Toronto, Canada

Abstract

Standard MPPCA denoising applies a fixed patch size uniformly across the entire brain — a fundamental mismatch with the structural heterogeneity of in-vivo tissue that blurs boundaries where precision matters most. We propose ak-MPPCA, which clusters voxel-wise gradient magnitudes via k-means and monotonically assigns discrete odd-integer kernel sizes, generating a per-voxel kernel map that autonomously adapts MPPCA denoising to local anatomical complexity — evaluated on HCP (MGH-USC) data at b-values of 5000–10000 s/mm².

Motivation

High b-value DWI (b ≥ 5000 s/mm²) is essential for probing restricted diffusion in tumour microstructure and axonal diameter estimation — but SNR degrades catastrophically, making denoising a clinical prerequisite.
Fixed-kernel MPPCA is fundamentally at odds with structural heterogeneity: a large kernel suppresses noise in homogeneous white matter but simultaneously blurs sharp tumour–parenchyma boundaries critical for surgical planning.
Existing adaptive methods (DESIGNER's adaptive patching) select voxels by spatial proximity and signal similarity — but do not leverage gradient-based structural information, leaving boundary-preserving denoising unsolved.
A voxel-wise kernel strategy that reads local gradient complexity enables simultaneous aggressive denoising in homogeneous regions and conservative denoising at boundaries — without manual parameter tuning or retraining.

Framework

ak-MPPCA Pipeline — Structure-aware Adaptive Kernel MPPCA

Pipeline: High b-value DWI stack → trace image at lowest b ≥ 100 → Sobel gradient magnitude with Gaussian smoothing (σ=4) → k-means clustering of voxel-wise gradient values into K bins → monotonic kernel assignment (high gradient → small kernel k_l=5, low gradient → large kernel k_u=9) → per-voxel odd-integer kernel size map → spatially-adaptive MPPCA denoising applied per-voxel → denoised DWI + colour FA maps.

Contributions

Gradient-based structural prior: Sobel edge detection on the trace image from the lowest available b-value (b ≥ 100) quantifies local tissue complexity, providing a hyperparameter-free, segmentation-free signal to drive per-voxel kernel selection.
Gradient-to-kernel mapping algorithm: k-means clustering of brain-masked gradient magnitudes into K discrete bins with monotonic assignment to odd-integer MPPCA kernel sizes — high gradient → small kernel (preserves boundaries), low gradient → large kernel (maximises SNR gain). Voxels outside the brain mask assigned k_u.
ak-MPPCA: full integration of the per-voxel kernel map into the MPPCA pipeline, maintaining all random-matrix-theory noise estimation and eigenvalue thresholding principles from Veraart et al. while replacing the fixed patch with an adaptive one.
Evaluated on 3 HCP (MGH-USC) subjects at b = 5000 and 10000 s/mm²; outperforms NYU-MPPCA, DIPY-MPPCA (kernel sizes 5, 7, 9), and DESIGNER adaptive patching in both DWI visual clarity and colour FA map accuracy — without any kernel size hyperparameter.

Results

Evaluated on MGH-1005, MGH-1006, and MGH-1010 from the HCP MGH-USC dataset at b = 5000 and 10000 s/mm². The proposed method consistently preserves structural detail in high-gradient boundary regions while achieving aggressive noise suppression in homogeneous tissue — outperforming NYU-MPPCA, DIPY-MPPCA, and adaptive patching across all subjects. Voxel-wise kernel maps confirm adaptive behaviour: smaller kernels at grey-white matter boundaries and cortical folds, larger kernels in deep white matter.

MGH-1010 (b=10000 s/mm²): (a) noisy DWI · (b) ak-MPPCA proposed · (c) adaptive patching · (d–f) DIPY-MPPCA at 5/7/9 · (g–i) NYU-MPPCA at 5/7/9 · (j) trace image · (k) voxel-wise kernel map. Yellow arrows show structural detail preserved by proposed method.

MGH-1006 (b=10000 s/mm²): Highlighted regions (yellow boxes) show where ak-MPPCA suppresses noise without blurring the boundary structure that fixed-size methods lose at kernel size 9.

MGH-1005 (b=5000 s/mm²): At lower b-value, ak-MPPCA maintains cleaner DWI than all baselines while the adaptive kernel map correctly identifies cortical boundary regions for conservative denoising.

Colour FA Maps

Primary diffusion directions encoded as RGB directionally encoded colour (DEC) maps for MGH-1005 across three slices. Top row: noisy. Bottom row: ak-MPPCA denoised. The proposed method restores directional coherence in noisy regions and sharpens crossing-fibre zones — validated by yellow arrows marking improved regions.

Slice 41: Noisy (top) vs. ak-MPPCA denoised (bottom). Directional coherence restored in crossing-fibre regions.

Slice 45: Yellow arrows mark regions where ak-MPPCA recovers clean diffusion orientation from heavily noisy DWI.

Slice 48: Deep white matter tracts show strong directional recovery after ak-MPPCA denoising.

BibTeX

@inproceedings{singhal2026embc, title = {Structure-aware Adaptive Kernel MPPCA Denoising for Diffusion MRI}, author = {Singhal, Ananya and Shanbhag, Dattesh Dayanand and Chatterjee, Sudhanya}, booktitle = {Engineering in Medicine and Biology Society (EMBC)}, address = {Toronto, Canada}, year = {2026} }