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

Sushant Agarwal, Sanjay Saxena, Alessandro Carriero, Gian Luca Chabert, Gobinath Ravindran, Sudip Paul, John R. Laird, Deepak Garg, Mostafa Fatemi, Lopamudra Mohanty, Arun Kumar Dubey, Rajesh Singh, Mostafa M. Fouda, Narpinder Singh, Subbaram Naidu, Klaudija Viskovic, Melita Kukuljan, Manudeep K. Kalra, Luca Saba, Jasjit S. SuriFrontiers Artif. Intell 2024 · DOI

TL;DR

This paper introduces COVLIAS 3.0, a cloud-based deep learning system that leverages quantized hybrid UNet3+ models for automated COVID-19 lesion detection in lung CT scans. By combining advanced deep learning architectures with quantization techniques, the system achieves improved accuracy and efficiency, addressing the limitations of manual analysis and previous AI models.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical need for rapid and accurate diagnosis, especially when RT-PCR tests are insufficient or delayed. While CT scans are valuable for detecting COVID-19 related lung lesions (ground-glass opacities, consolidations), manual analysis by radiologists is time-consuming and prone to fatigue. Previous deep learning approaches for this task, while promising, often suffered from moderate performance and lacked efficiency for widespread deployment. This paper aims to bridge this gap by developing a more accurate, efficient, and cloud-deployable AI system for COVID-19 lesion segmentation.

Methods

The core of COVLIAS 3.0 is a hybrid deep learning (HDL) framework built upon the UNet3+ architecture. The researchers developed two HDL models: VGG-UNet3+ and ResNet-UNet3+, which integrate established convolutional neural network backbones (VGGNet and ResNet, respectively) with the UNet3+ encoder-decoder structure. UNet3+ is notable for its "full-scale skip connections," which allow for the integration of low-level image features with high-level semantic information across different network layers, enhancing feature representation.

To address efficiency concerns, the models were subjected to quantization. This process reduces the precision of the model's weights and parameters, significantly decreasing model size and computational requirements without a substantial loss in accuracy. The study also implemented a cloud-based framework using Amazon Web Services for accessibility and scalability.

Two loss functions were employed during training: Dice Similarity (DS) and Binary Cross-Entropy (BCE). Performance was evaluated using metrics such as Area Error (AE), Dice Similarity (DS), Jaccard Index (JI), Bland-Altman plots, and Correlation Coefficient plots.

Datasets

The study utilized a dataset of approximately 3,500 CT images from 45 COVID-19 positive patients from the University Hospital for Infectious Diseases in Zagreb, Croatia, for training and validation. An additional 500 CT scans from the MosMed dataset (Russia) were used as an independent, unseen test set to evaluate the generalizability of the models. Annotations were provided by expert radiologists.

Key Insights and Ideas

  • The Power of Full-Scale Skip Connections: The UNet3+ architecture's "full-scale skip connections" are crucial for effectively combining fine-grained image details with broader contextual information, leading to improved lesion segmentation.
  • Hybrid Architectures Outperform Solo DL: Combining established CNN backbones (VGG, ResNet) with UNet3+ (HDL models) demonstrated superior performance compared to the standalone UNet3+ model (SDL). This suggests that leveraging pre-trained feature extraction capabilities alongside UNet3+'s architectural strengths is highly beneficial.
  • Quantization is Key for Deployment: Quantization significantly reduces model size and computational load, making these powerful deep learning models more practical for deployment in cloud environments and potentially on edge devices.
  • Cloud-Based AI for Accessibility: Deploying the system on a cloud platform (AWS) enhances accessibility for healthcare professionals, enabling wider adoption and faster integration into clinical workflows.

Key Results

  • The ResNet-UNet3+ model demonstrated superior performance, outperforming the UNet3+ model by 17% for Dice loss and 10% for BCE loss on the unseen dataset.
  • Quantization led to substantial model compression: UNet3+ achieved 66.76% reduction, VGG-UNet3+ 36.64%, and ResNet-UNet3+ 46.23%.
  • The quantized models achieved very fast prediction times, with ResNet-UNet3+ and VGG-UNet3+ taking less than 0.5 seconds per slice.
  • Statistical tests (Mann-Whitney, Paired t-Test, Wilcoxon, Friedman) confirmed the stability and reliability of the proposed system with p < 0.001.
  • The COVLIAS 3.0 system achieved a Dice Similarity Coefficient (DSC) that was 16% better than the mean DSC reported in previous studies.

Limitations and Future Work

  • Feature Extraction Limitations: The authors acknowledge that the current model might have limitations in extracting very small features. They suggest incorporating attention models or transformer models to enhance feature extraction capabilities.
  • Hardware Dependency: The current protocol requires hardware like GPUs for training, which might be a limitation for some institutions.
  • Future Work: The authors propose extending the framework to include more variants of hybrid systems (DL with ML), training on datasets from diverse ethnicities and countries to improve robustness, and exploring applications in other medical imaging domains (mammography, urology, etc.). They also suggest exploring empirical convergence techniques for training.

Summary updated

BibTeX
@article{agarwal2024covlias,
  author = {Sushant Agarwal and Sanjay Saxena and Alessandro Carriero and Gian Luca Chabert and Gobinath Ravindran and Sudip Paul and John R. Laird and Deepak Garg and Mostafa Fatemi and Lopamudra Mohanty and Arun Kumar Dubey and Rajesh Singh and Mostafa M. Fouda and Narpinder Singh and Subbaram Naidu and Klaudija Viskovic and Melita Kukuljan and Manudeep K. Kalra and Luca Saba and Jasjit S. Suri},
  title = {{COVLIAS 3.0: cloud-based quantized hybrid UNet3+ deep learning for COVID-19 lesion detection in lung computed tomography}},
  journal = {Frontiers Artif. Intell.},
  year = {2024},
  doi = {10.3389/frai.2024.1304483},
}
Figure