Soft and self constrained clustering for group-based labeling.

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Biology

Background:

• Deep neural networks require large labeled medical image datasets, demanding significant expert effort.
• Group-based labeling, involving clustering and labeling image groups, can reduce costs but relies on cluster purity.

Purpose of the Study:

• To address challenges in conventional constrained clustering for medical images, specifically inappropriate constraints and the effort required for constraint definition.
• To propose and evaluate novel soft-constrained and self-constrained clustering methods.

Main Methods:

• A soft-constrained clustering approach was developed to disregard unsuitable constraints, mitigating the semantic-visual similarity gap.
• A self-constrained clustering method was introduced, leveraging prior image knowledge for automatic constraint generation, reducing expert input.

Main Results:

• Experiments on endoscopic image datasets demonstrated that the proposed soft-constrained and self-constrained clustering methods significantly improved cluster purity.
• The novel methods effectively handled inappropriate constraints and reduced the need for manual constraint input from medical experts.

Conclusions:

• The developed soft-constrained and self-constrained clustering methods offer a more efficient and accurate approach to medical image clustering.
• These methods have the potential to lower the cost and effort associated with preparing labeled datasets for deep neural network applications in medical imaging.