COVLIAS 3.0: cloud-based quantized hybrid UNet3+ deep learning for COVID-19 lesion detection in lung computed...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed Tomography (CT) scans are crucial for COVID-19 diagnosis when RT-PCR is insufficient, particularly for identifying lung opacities.
• Manual lesion detection in CT scans is challenging for radiologists.
• Previous solo deep learning (SDL) models showed limited performance in COVID-19 detection.

Purpose of the Study:

• To develop and evaluate novel cloud-based quantized hybrid deep learning (HDL) models for enhanced COVID-19 lesion detection in CT scans.
• To improve upon the performance of solo deep learning (UNet3+) models.

Main Methods:

• Trained one SDL (UNet3+) and two HDL models (VGG-UNet3+, ResNet-UNet3+) using expert radiologist annotations on 3,500 CT scans.
• Employed 5-fold cross-validation and tested on 500 unseen CT scans within a cloud framework.
• Utilized Dice Similarity (DS) and binary cross-entropy (BCE) loss functions for training and evaluated performance using various metrics.

Main Results:

• The ResNet-UNet3+ HDL model demonstrated superior performance, outperforming UNet3+ by 17% for Dice loss and 10% for BCE loss.
• Quantization reduced model sizes significantly: UNet3+ by 66.76%, VGG-UNet3+ by 36.64%, and ResNet-UNet3+ by 46.23%.
• Statistical tests confirmed the stability and reliability of the models (p < 0.001).

Conclusions:

• Hybrid deep learning models incorporating full-scale skip connections with VGG and ResNet architectures significantly improve COVID-19 detection accuracy in CT scans.
• The developed HDL models offer a more effective and efficient solution for radiologists in diagnosing COVID-19.
• Quantized HDL models provide a balance of performance and reduced computational footprint.