Fine-Tuning Deep Learning by Crowd Participation.

Area of Science:

• Medical image analysis
• Deep learning
• Computational pathology

Background:

• Limited annotated data is a major challenge for deep learning (DL) in medical image analysis.
• Collecting high-quality annotations is time-consuming, costly, and requires domain expertise.
• This hinders the development of large-scale medical imaging databases.

Purpose of the Study:

• To address the challenge of limited annotated data in medical image analysis.
• To develop novel methods for training DL models using crowdsourced and less domain-specific data.
• To improve the efficiency and accessibility of medical image annotation for DL.

Main Methods:

• Introduced a robust aggregation layer for convolutional neural networks to handle noisy crowdsourced annotations [S5].
• Developed a method to translate biomedical images into star-shaped objects embeddable in game canvases, reducing annotation domain knowledge requirements [S6].
• Validated methods on breast cancer histology images and compared performance against baseline approaches.

Main Results:

• The robust aggregation method showed significant improvements over majority voting for noisy annotations.
• The image-to-game object translation technique demonstrated promising results compared to conventional crowdsourcing.
• Both methods effectively reduced the reliance on extensive domain expertise for data annotation.

Conclusions:

• Novel DL approaches can effectively utilize noisy crowdsourced data for medical image analysis.
• Transforming images into game-like objects offers a promising alternative to traditional annotation methods.
• These advancements facilitate the development of DL models with less annotated data and reduced expert input.