Learning from crowds in digital pathology using scalable variational Gaussian processes.

Area of Science:

• Computational pathology
• Machine learning for medical imaging

Background:

• Scaling data labeling is crucial for machine learning, especially in expert-intensive fields like pathology.
• Crowdsourcing offers a solution, but learning from noisy, crowd-sourced labels presents challenges.

Purpose of the Study:

• To investigate the efficacy of Scalable Variational Gaussian Processes for Crowdsourcing (SVGPCR) in digital pathology.
• To compare SVGPCR performance against other crowdsourcing methods and gold-standard labels.

Main Methods:

• Application of SVGPCR to a large multi-rater dataset of breast cancer tissue annotations.
• Comparison with deep learning-based crowdsourcing methods and pathologist-generated labels.
• Analysis of SVGPCR's ability to estimate annotator reliabilities.

Main Results:

• SVGPCR demonstrated competitive performance against gold-standard pathologist labels.
• SVGPCR met or exceeded the performance of other deep learning-based crowdsourcing methods.
• The method effectively learned class-conditional reliabilities of individual annotators.

Conclusions:

• SVGPCR can successfully engage non-experts in pathology labeling tasks.
• Estimated annotator reliabilities can guide task assignment for improved efficiency.
• Gaussian-process classifiers show comparable performance to deep learning methods in this context.