A Few-Shot U-Net Deep Learning Model for COVID-19 Infected Area Segmentation in CT Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Radiographic patterns on CT chest scans show high sensitivity and specificity for COVID-19 identification.
• Traditional supervised learning for CT image segmentation requires extensive data and static network weights.
• Few-shot learning (FSL) offers a solution for training models with limited data.

Purpose of the Study:

• To evaluate the effectiveness of few-shot learning (FSL) within U-Net architectures for segmenting pneumonia-infected areas in CT images for COVID-19 detection.
• To investigate if dynamic fine-tuning of network weights using FSL improves segmentation accuracy compared to traditional methods.

Main Methods:

• Implementation of few-shot learning (FSL) principles within U-Net deep learning architectures.
• Dynamic fine-tuning of U-Net network weights as new, limited samples are introduced.
• Utilizing 4-fold cross-validation to assess segmentation performance using metrics like Intersection over Union (IoU) and F1 score.

Main Results:

• The proposed few-shot U-Net architecture demonstrated improved segmentation accuracy for COVID-19 infected regions.
• An average improvement of 5.388 ± 3.046% in IoU and 5.394 ± 3.015% in F1 score was observed across all test data.
• The Kruskal-Wallis test confirmed the statistical significance of the improvement (p-value = 0.026) compared to traditional U-Net models.

Conclusions:

• Few-shot learning integrated into U-Net architectures significantly enhances the accuracy of segmenting COVID-19 related pneumonia in CT images.
• This FSL approach offers a more efficient method for COVID-19 detection using CT scans, particularly when large datasets are unavailable.
• The dynamic fine-tuning capability of the proposed model represents a advancement in medical image analysis for infectious disease identification.