Task description¶

Task : SSL for 3D LSM image segmentation under fixed training data conditions¶

Participants are restricted to using the provided, processed image data for methodology development. Based on insights from the 2025 edition, this year we will combine all structure types within a single training set to fully explore the potential of SSL strategies for robust representation learning across diverse biological structures. However, unlike the 2025 edition,Task 1 will provide cropped, unannotated patches for SSL rather than the original raw 3D LSM images.

In addition, this year we categorize biological structures into four groups based on both morphological characteristics and spatial density: isolated sparse, isolated dense, contiguous sparse, and contiguous dense. Model performance will then be assessed separately within each category, enabling a more refined evaluation of model generalization and clearer identification of model strengths and limitations across varying structural complexities.

For Task 1, participants will be provided with a training dataset consisting of two subsets:

1. Unannotated subset: a large collection of patches cropped and preprocessed from 3D LSM images containing isolated sparse, isolated dense, contiguous sparse, and contiguous dense structures derived from multiple specimens. This subset includes more than 35,050 patches of size 300×300×300 voxels. These unannotated patches are intended for SSL-based model pretraining, allowing models to learn generalizable representations across diverse biological structures.

2. Annotated subset: a curated selection of 3D LSM image patches representing the same four structure categories included in the unannotated subset, accompanied by precise manual annotations. This subset contains more than 210 patches with manual voxel-wise annotations and enables participants to fine-tune their pretrained models for general segmentation of diverse biological structures.

Task : Self-supervised learning for 3D LSM image segmentation under open data conditions¶

While Task 1 focuses on self-supervised learning (SSL) strategies for 3D LSM image segmentation under fixed training data conditions, Task 2 is designed to investigate the effect of data scaling when applying SSL to the light-sheet microscopy domain. In contrast to Task 1, participants in Task 2 are allowed to leverage arbitrary amounts of data for SSL.

We will provide access to the full set of original unannotated 3D LSM images used in Task 1, prior to any cropping or preprocessing. Participants may therefore develop and apply their own preprocessing or curation pipelines to construct pretraining datasets for SSL. In addition, participants are encouraged to incorporate any additional public or private LSM data.

The finetuning stage is constrained to use the same annotated training set as Task 1 , without any external annotated data.This constraint ensures that the performance gains can be attributed to data scaling in SSL rather than to additional supervision from extra annotations.