Leaderboard 2: 3D Denoising (Tribolium)¶

👉 Download the training data here - Data on Zenodo

General Information¶

This leaderboard focuses on denoising 3D fluorescence microscopy images of Tribolium castaneum (the red flour beetle), a classic model for studying early embryo development.

The dataset comes from the Tribolium denoising dataset originally published in Weigert, M., Schmidt, U., Boothe, T. et al. Content-aware image restoration: pushing the limits of fluorescence microscopy. Nat Methods 15, 1090–1097 (2018).

The dataset features fluorescently labeled nuclei imaged using a multiphoton laser-scanning microscope. Each embryo was imaged under four different laser power levels, producing 3D volumes with varying levels of noise, mimicking the challenges of live-cell imaging under low-light conditions.

To prepare the dataset for supervised training, the raw large 3D volumes were cropped into stacks of size (16, 64, 64). For each sample, we provide:

1. A noisy input volume (from one of three lower-exposure conditions)
2. A corresponding ground truth (GT) volume (from the highest-quality acquisition using higher laser power and longer exposure)

Training Dataset¶

The training dataset contains noisy-GT image pairs, extracted as small 3D crops from larger embryo volumes. These local patches cover diverse spatial regions and imaging conditions.

• All stacks are of shape (16, 64, 64)
• Contains 15,500 paired samples
• Files are aligned across folders, e.g., noisy/00023.tif ↔ gt/00023.tif

These cropped volumes provide sufficient variation in noise, structure to train supervised 3D denoising models.

Evaluation¶

For evaluation, models will be tested on full-sized 3D embryo stacks (average shape ~40×600×600) that were not included in the training data. These held-out volumes represent different developmental stages, allowing us to assess model generalization to new morphological contexts and time points.

Data samples used for the evaluation might look like this:

Your goal is to build a model that restores these noisy volumes while preserving nuclear structure and spatial accuracy.

👀 Check out the example submission here - Example submission on GitHub

✍️ Submit here - Submission page on GC