Deep learning is combined with massive-scale citizen science to improve large-scale image classification.

Area of Science:

• Life Sciences
• Bioinformatics
• Cell Biology

Background:

• Accurate pattern recognition and classification of microscopy images are critical in life sciences.
• Large-scale classification of subcellular localization patterns presents a significant challenge.

Purpose of the Study:

• To combine human intelligence and deep learning for enhanced large-scale fluorescence microscopy image classification.
• To develop an automated tool for classifying protein subcellular localization patterns.

Main Methods:

• Integrated an image-classification mini-game (Project Discovery) into EVE Online using the Human Protein Atlas (HPA) Cell Atlas dataset.
• Developed a deep learning model (Localization Cellular Annotation Tool - Loc-CAT) for classifying 29 subcellular localization patterns.
• Applied transfer learning to combine gamer annotations and deep learning for a boosted classification model.

Main Results:

• Over 33 million classifications were provided by 322,000 gamers, identifying novel localization patterns.
• The Loc-CAT tool efficiently classified proteins, including multi-localization ones, across cell types.
• The combined gamer-AI approach achieved an F1 score of 0.72 for characterizing subcellular protein distribution.

Conclusions:

• Engaging commercial computer game players provides valuable data to augment deep learning.
• This hybrid approach enables scalable and improvable large-scale image classification for biological research.
• Crowdsourcing via gaming significantly enhances the accuracy and scope of cell image analysis.