Toward data-efficient learning: A benchmark for COVID-19 CT lung and infection segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate segmentation of lung and infection in COVID-19 CT scans is crucial for patient management.
• Existing segmentation methods rely on large, private datasets, hindering development and comparison.
• Radiologists' workload during the pandemic limits the creation of extensive annotated datasets.

Purpose of the Study:

• To establish data-efficient deep learning benchmarks for COVID-19 CT segmentation.
• To facilitate fair comparison of segmentation methods across different datasets and settings.
• To promote the development of deep learning models using limited annotated data.

Main Methods:

• Developed three benchmarks for lung and infection segmentation using 70 annotated COVID-19 cases.
• Incorporated active research areas like few-shot learning, domain generalization, and knowledge transfer.
• Provided standard training/validation/testing splits, evaluation metrics, and code for reproducible research.

Main Results:

• Offered over 40 pretrained baseline models based on state-of-the-art networks.
• Achieved high Dice Similarity Coefficient (DSC) scores: 97.3% (left lung), 97.7% (right lung), 67.3% (infection).
• Attained excellent Normalized Surface Dice (NSD) scores: 90.6% (left lung), 91.4% (right lung), 70.0% (infection).

Conclusions:

• Presents the first data-efficient learning benchmark for medical image segmentation.
• Provides the largest collection of pretrained models for COVID-19 CT segmentation to date.
• Publicly available resources aim to accelerate deep learning research for efficient segmentation with limited data.