Segmenting functional tissue units across human organs using community-driven development of generalizable machine...

Area of Science:

• Anatomical imaging and bioinformatics
• Computational biology and machine learning
• Digital pathology and histology

Background:

• Automated segmentation of anatomical structures is crucial for building a reference atlas of the healthy human body.
• Bioimages from diverse sources present challenges due to variations in sample preparation.
• The Human Biomolecular Atlas (HuBMAP) and Human Protein Atlas (HPA) initiatives aim to create comprehensive atlases.

Approach:

• A machine learning algorithm development competition, "Hacking the Human Body," was hosted on Kaggle.
• 1,175 teams from 78 countries participated in open-science code development.
• Histology images from two consortia across five organs were used for model training and validation.

Key Points:

• The competition generated machine learning models capable of segmenting anatomical structures in histology images.
• Benchmark datasets and insights into participant challenges were documented.
• The developed models will be integrated into the HuBMAP data portal for large-scale data processing.

Conclusions:

• Community-driven, open-science efforts can accelerate the development of advanced bioimage analysis tools.
• The competition successfully fostered innovation in automated anatomical segmentation.
• The resulting models will significantly support the construction of the Human Reference Atlas.