In the past decade, we have seen an increasing number of large-scale bioimage initiatives. These efforts have generated a rich source of histology images, covering a plethora of organs and tissues in the human body, which should ultimately come together to form a reference map of the human body.

Combining massive amounts of imaging data requires harmonization and consensus image analysis pipelines, including segmentation algorithms that perform well on images from diverse sources. This led researchers from the Human Protein Atlas (HPA) and the Human Biomolecular Atlas (HuBMAP) to co-host a community-driven machine learning challenge, called ?Hacking the Human Body ?, on the Kaggle platform. The challenge setup and the results were recently presented in Nature Communications.

The competition engaged 1175 teams, bringing together people with various expertise and from 78 countries, who not only competed, but also collaborated and interacted extensively. The challenge focused on segmentation of Functional Tissue Units (FTUs) in five different organs using tissue images from both the HPA and HuBMAP. An FTU is defined as the smallest anatomical structure that performs a unique physiologic function in an organ, such as alveoli in the lung or glomeruli in the kidney. As their structure and composition are often subjected to alterations in human diseases, robust segmentation of FTUs is an important step towards image segmentation tasks in medical settings. The code from the winning models will be productized and deployed in the HuBMAP, but all data and code is also publicly available at GitHub and Zenodo.

Link to article