Comparing Different Data Partitioning Strategies for Segmenting Areas Affected by COVID-19 in CT Scans.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Diseases

Background:

• Reverse Transcription Polymerase Chain Reaction (RT-PCR) is the gold standard for COVID-19 diagnosis.
• Computed tomography (CT) imaging is crucial for confirming diagnoses in symptomatic patients with negative RT-PCR results.
• Accurate segmentation of COVID-19 affected lung areas in CT scans is essential for diagnosis and monitoring.

Purpose of the Study:

• To evaluate automatic segmentation methods for COVID-19 areas in CT scans using deep neural networks.
• To compare the performance of automatic segmentation with manual segmentation and interobserver agreement among radiologists.
• To assess the impact of different data partitioning strategies (CT-scan vs. slice) on segmentation accuracy.

Main Methods:

• Utilized two datasets and two deep neural network architectures, including U-Net and a novel architecture.
• Implemented both manual and automatic lung segmentation techniques.
• Evaluated segmentation performance using Dice metrics under CT-scan and slice strategies.

Main Results:

• Automatic segmentation achieved Dice metrics of 73.01% (CT-scan strategy) and 84.66% (slice strategy) with automatic lung segmentation.
• With manual lung segmentation, Dice metrics for COVID-19 segmentation were 74.47% (CT-scan) and 85.35% (slice).
• No statistically significant difference was found between automatic segmentation performance and interobserver agreement in a subset of 7 CT scans.

Conclusions:

• The slice strategy demonstrated slightly superior performance for COVID-19 segmentation compared to the CT-scan strategy.
• Automatic segmentation performance closely matched interobserver agreement among radiologists, suggesting its clinical utility.
• Deep learning models show promise for automated analysis of COVID-19 in CT imaging.