ENRICHing medical imaging training sets enables more efficient machine learning.

Area of Science:

• Biomedical imaging
• Machine learning
• Data science

Background:

• Deep learning (DL) models require extensive labeled data for training and testing in biomedical imaging.
• Human expert labeling is a bottleneck due to limitations in availability and the computational expense of processing large datasets.
• Instance selection is crucial for prioritizing images that most effectively improve DL model performance.

Purpose of the Study:

• To introduce ENRICH (Eliminate Noise and Redundancy for Imaging Challenges), a novel method for prioritizing images in DL training datasets.
• To address the challenge of selecting the most informative images to enhance DL model performance while minimizing computational costs and labeling efforts.
• To develop a customizable approach for instance selection tailored to the unique characteristics of medical image datasets.

Main Methods:

• ENRICH prioritizes images based on the diversity they contribute to the training set.
• The method was evaluated on classification and segmentation tasks across multiple medical image datasets.
• Performance was compared against random image selection as a control.

Main Results:

• Medical image datasets generally exhibit lower inter-image diversity compared to non-medical datasets.
• ENRICH achieved near-maximal performance on classification and segmentation tasks using a fraction of the available images, without requiring upfront labeling.
• ENRICH demonstrated superior performance over random image selection and proved effective in identifying dataset errors and outliers.

Conclusions:

• ENRICH offers a computationally efficient and straightforward method for prioritizing images for expert labeling in DL applications.
• The approach optimizes the use of limited resources by focusing on the most valuable data for training DL models.
• ENRICH facilitates improved efficiency and accuracy in developing DL models for biomedical imaging.