AI4Life Denoising Challenge 2024: Results

by Vera Galinova, Beatriz Serrano-Solano

The AI4Life Denoising Challenge, which aimed to improve the denoising of microscopy images using deep learning, has been successfully completed. With 104 registrations from 27 countries, the challenge gathered participation from the global scientific community, generating 151 total submissions across four distinct leaderboards featuring different datasets containing one of two noise types.

What was the challenge about?

Microscopy images are indispensable for biological and medical research, but noise introduced during acquisition can impair image quality and complicate interpretation. The challenge’s focus was the unsupervised denoising task, which, unlike supervised learning, does not require pairs of noisy and clean images, simplifying the real-world application of the algorithms. Researchers applied advanced deep learning algorithms to real-world datasets, aiming to improve image quality while preserving essential features such as edges, textures, and fine details. Find all the information on the challenge at https://ai4life.eurobioimaging.eu/denoising-challenge/

Key statistics

Total Participants: 104 registrants from 27 countries
Total Submissions: 151
Leaderboard Submissions:
Structured Noise 1: 17 submissions from 7 participants
Structured Noise 2: 7 submissions from 5 participants
Unstructured Noise 1: 16 submissions from 7 participants
Unstructured Noise 2: 13 submissions from 8 participants

Results by category

Structured Noise 1: Hagen et al (50 total submissions)

	Name	Affiliation	Algorithm	Result SI-PSNR	Result SSIM	Code
1	bensalmon	University of Birmingham	COSDD	31.4847	0.5523	Link to code
2	mcroft	Human Technopole	N2V	30.8962	0.5534	Link to code
3	shanetoy	Seoul National University	N2V	30.8406	0.5525	Link to code

Structured Noise 2: SUPPORT (20 total submissions)

	Name	Affiliation	Algorithm	Result SI-PSNR	Result SSIM	Code
1	bensalmon	University of Birmingham	COSDD	30.4415	0.6278	Link to code
2	a897574323	University of Shanghai for Science and Technology	N2V	29.7744	0.6112	Link to code
3	edoardogiacomello	Human Technopole	N2V	28.8547	0.5272	Link to code

Unstructured Noise 1: JUMP (47 total submissions)

	Name	Affiliation	Algorithm	Result SI-PSNR	Result SSIM	Code
1	bensalmon	University of Birmingham	COSDD	35.6282	0.9507	Link to code
2	mcroft	Human Technopole	N2V	35.4957	0.9413	Link to code
3	edoardogiacomello	Human Technopole	N2V	35.4957	0.9412	Link to code

Unstructured Noise 2: W2S (34 total submissions)

	Name	Affiliation	Algorithm	Result SI-PSNR	Result SSIM	Code
1	bensalmon	University of Birmingham	COSDD	35.6855	0.9163	Link to code
2	edoardogiacomello	Human Technopole	N2V2	35.0505	0.9025	Link to code
3	edoardogiacomello	Human Technopole	N2V	35.0319	0.9027	Link to code

What’s next?

The AI4Life Denoising Challenge provided a platform for benchmarking denoising methods across various datasets. The challenge showed how unsupervised deep learning algorithms, like COSDD and N2V, can significantly enhance image quality. The leaderboard results asserted the robustness of these methods across different datasets and noise types.

We encourage researchers to continue to participate in the challenge through late submissions. Please contact us if you want to participate by submitting your algorithm, and the challenge team will assist you.

Looking ahead, AI4Life plans to host more challenges soon. Stay tuned for upcoming announcements!