Hyperparameter optimization for image analysis: application to prostate tissue images and live cell data of...

Area of Science:

• Computational Biology
• Image Analysis
• Machine Learning

Background:

• Automated microscopy image analysis relies on complex pipelines with numerous methods.
• Optimizing method-dependent hyperparameters is crucial for pipeline performance.
• Gradient calculation for loss functions in these pipelines is often infeasible, preventing traditional optimization methods.

Purpose of the Study:

• To introduce HyperHyper, a novel framework for zero-order black-box hyperparameter optimization.
• To enable modular separation of hyperparameter sampling and optimization strategies.
• To develop a loss function visualization technique using infimum projection for enhanced insight.

Main Methods:

• Developed the HyperHyper framework with a modular architecture.
• Implemented separation of hyperparameter sampling and optimization.
• Utilized infimum projection for loss function visualization.

Main Results:

• Applied HyperHyper across three experiments with diverse imaging modalities, evaluating over 400,000 hyperparameter combinations.
• Optimized pipelines for cell nuclei segmentation in prostate tissue and hepatitis C virus protein detection in live cells.
• Demonstrated improved results by separating sampling and optimization strategies and using infimum projection for visualization.

Conclusions:

• The modular HyperHyper framework, separating sampling and optimization, significantly improves performance in microscopy image analysis pipelines.
• Infimum projection-based visualization offers valuable insights into the hyperparameter optimization process.