US Speckle Reduction Part 2 — Classical vs DnCNN

Speckle Reduction on Ultrasound Images — Classical vs DnCNN

Background

My PhD research focused on freehand B-mode ultrasound image quality —
speckle noise, spatial correlation modeling, and super-resolution restoration.
That was 8 years ago. This project is me rebuilding that work with modern tools.

Starting point: a clean benchmark on the
BUSI dataset
(780 breast ultrasound images — benign, malignant, normal).

The Problem with Standard Benchmarks

Most speckle reduction papers evaluate on the full image — background included.
But in clinical practice, what matters is image quality within the lesion.

So I added ROI-based evaluation using the segmentation masks provided in BUSI.
For images with multiple lesions, masks were combined via union.

def evaluate_with_roi(original, denoised, mask=None):
    global_psnr = psnr(original, denoised, data_range=1.0)
    global_ssim = ssim(original, denoised, data_range=1.0)

    if mask is not None and mask.max() > 0:
        # ROI bounding box
        rows = np.any(mask > 0, axis=1)
        cols = np.any(mask > 0, axis=0)
        rmin, rmax = np.where(rows)[0][[0, -1]]
        cmin, cmax = np.where(cols)[0][[0, -1]]

        if (rmax - rmin) >= 7 and (cmax - cmin) >= 7:
            roi_orig = original[rmin:rmax, cmin:cmax]
            roi_den  = denoised[rmin:rmax, cmin:cmax]
            roi_ssim = ssim(roi_orig, roi_den, data_range=1.0)

Methods

Noise model: Additive Gaussian (σ=0.05) as speckle proxy (Note: not physically accurate — Nakagami or correlated models are more realistic. PSF-aware correlated noise model is next on the list.)
Non-Local Means (NLM)
Median Filter
Gaussian Filter
DnCNN — 17-layer residual CNN, trained from scratch on BUSI

DnCNN — A Note on Training

Training was not straightforward.

Initial run with lr=1e-3 plateaued immediately after epoch 2:
Epoch [ 2/30] — Loss: 0.002472

Epoch [10/30] — Loss: 0.002408

Epoch [20/30] — Loss: 0.002383

Epoch [30/30] — Loss: 0.002363 ← barely moved

Dropping to lr=1e-4 with step decay fixed it:
Epoch [ 5/50] — Loss: 0.001487

Epoch [20/50] — Loss: 0.001023

Epoch [50/50] — Loss: 0.000899

This is worth noting: DnCNN is sensitive to initialization and learning rate.
Results varied across runs (PSNR range ~31.2–31.7 dB). Single-run DL benchmarks
should be interpreted with some caution.

Results (780 images)

Method	PSNR Global	SSIM Global	PSNR ROI	SSIM ROI
Noisy	26.25	0.649	26.34	0.681
Median Filter	29.71	0.810	29.32	0.811
Gaussian Filter	29.99	0.836	29.76	0.833
NLM	31.47	0.827	31.52	0.842
DnCNN	31.24	0.847	31.17	0.860

Key Findings

1. NLM vs DnCNN — it depends on the metric.
NLM wins on PSNR, DnCNN wins on SSIM. SSIM captures structural similarity
better than pixel-level fidelity — arguably more relevant for diagnostic imaging.

2. Global metrics can mislead.
Median filter looks decent globally (29.71 dB) but drops in ROI (29.32 dB) —
edge blurring at lesion boundaries. A paper reporting only global metrics
would miss this.

3. Dataset-scale evaluation matters.
On individual images, NLM sometimes outperforms DnCNN. The pattern only
stabilizes at dataset scale. Single-image comparisons in this field are unreliable.

What's Next

Super-resolution: Real-ESRGAN / SwinIR fine-tuned on ultrasound
Realistic noise model: PSF-aware correlated speckle from phantom data (I actually collected this data during my PhD — time to revisit it)
Thyroid nodule segmentation with SAM

Code on GitHub 👇

https://github.com/minebyte/us-speckle-reduction-benchmark

DEV Community