[Paper Review] Automating Parameter Selection in Deep Image Prior for Fluorescence Microscopy Image Denoising via Similarity-Based Parameter Transfer

Lina Meyer, Felix Wissel+9 more|arXiv (Cornell University)|Jan 17, 2026

Cell Image Analysis Techniques0 citations

TL;DR

AUTO-DIP transfers DIP denoising parameters from a calibration set to new fluorescence microscopy images using similarity-based strategies, achieving faster and improved denoising compared to the original DIP and a variational method.

ABSTRACT

Unsupervised deep image prior (DIP) addresses shortcomings of training data requirements and limited generalization associated with supervised deep learning. The performance of DIP depends on the network architecture and the stopping point of its iterative process. Optimizing these parameters for a new image requires time, restricting DIP application in domains where many images need to be processed. Focusing on fluorescence microscopy data, we hypothesize that similar images share comparable optimal parameter configurations for DIP-based denoising, potentially enabling optimization-free DIP for fluorescence microscopy. We generated a calibration (n=110) and validation set (n=55) of semantically different images from an open-source dataset for a network architecture search targeted towards ideal U-net architectures and stopping points. The calibration set represented our transfer basis. The validation set enabled the assessment of which image similarity criterion yields the best results. We then implemented AUTO-DIP, a pipeline for automatic parameter transfer, and compared it to the originally published DIP configuration (baseline) and a state-of-the-art image-specific variational denoising approach. We show that a parameter transfer from the calibration dataset to a test image based on only image metadata similarity (e.g., microscope type, imaged specimen) leads to similar and better performance than a transfer based on quantitative image similarity measures. AUTO-DIP outperforms the baseline DIP (DIP with original DIP parameters) as well as the variational denoising approaches for several open-source test datasets of varying complexity, particularly for very noisy inputs. Applications to locally acquired fluorescence microscopy images further proved superiority of AUTO-DIP.

Motivation & Objective

• Address the need for automatic, image-specific DIP parameter selection in fluorescence microscopy denoising to improve speed and quality.
• Investigate whether parameter transfer based on image similarity can replace per-image grid searches.
• Identify the most effective similarity criteria for parameter transfer across microscope types, specimens, and imaging modalities.
• Evaluate AUTO-DIP against the original DIP configuration and a state-of-the-art variational denoising method across diverse fluorescence datasets.

Proposed method

• Use an unsupervised Deep Image Prior (DIP) framework with a U-net backbone for image denoising.
• Define a grid search over DIP architecture (depth, width, skip connections) and stopping iterations to establish an optimal configuration per calibration image.
• Construct a calibration dataset (n=110) and a validation set (n=55) from the FMD dataset to explore parameter transfer strategies.
• Evaluate three similarity criteria for parameter transfer: Group-Based (microscope-specimen groups), Metric-Based (nearest neighbor in pixel/perceptual space), and a Combined Group-Metric variant.
• Compare transfer strategies to image-specific optimal configurations, the original DIP parameters, and a sparsity-based variational denoising method.
• Test generalizability on additional public datasets (Hagen, BioSR, W2S Shah, and internal UKE data).

Experimental results

Research questions

• RQ1 Can DIP hyperparameters (architecture and stopping point) be effectively transferred between similar fluorescence microscopy images to achieve image-specific denoising without per-image optimization?
• RQ2 Which similarity criteria (metadata-based groups vs. pixel/semantic similarity) best guide parameter transfer for DIP in fluorescence microscopy?
• RQ3 How does AUTO-DIP perform relative to the original DIP configuration and a state-of-the-art variational denoising method across multiple datasets and noise levels?

Key findings

• Parameter transfer based on microscope-specimen group similarity yields the best average denoising performance (mean PSNR 34.99, LPIPS 0.098) on the validation set, approaching the image-specific optimum (35.56 PSNR).
• Across evaluated metrics, AUTO-DIP generally outperforms the original DIP configuration (average PSNR improvement from 31.90 to higher values) and the variational sparsity-based method in many test cases.
• AUTO-DIP denoising runs ~3x faster on average than the original DIP configuration.
• AUTO-DIP shows robust improvements for very noisy inputs and provides better denoising across diverse fluorescence microscopy datasets compared to baseline and some competing methods.
• Not all cases benefit equally; some fine structural details may be slightly blurred or oversmoothed in a few instances, depending on dataset and noise level.