Related Experiment Video
Updated: Oct 25, 2025
Lung CT Segmentation to Identify Consolidations and Ground Glass Areas for Quantitative Assesment of SARS-CoV Pneumonia
Published on: December 19, 2020
CoLe-CNN+: Context learning - Convolutional neural network for COVID-19-Ground-Glass-Opacities detection and
Giuseppe Pezzano1, Oliver Díaz2, Vicent Ribas Ripoll3
1Eurecat, Centre Tecnològic de Catalunya, EHealth Unit, Barcelona, Spain; Universitat de Barcelona, Department of Mathematics and Computer Science, Barcelona, Spain.
This study introduces an automated deep learning system for COVID-19 detection and lesion segmentation using CT scans, offering higher accuracy than traditional PCR tests. The AI system achieves superior performance in identifying and segmenting lung infections, aiding in more precise COVID-19 diagnosis.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Deep Learning for Disease Detection
- Radiology and Pulmonary Medicine
Background:
- Reverse Transcription-Polymerase Chain Reaction (RT-PCR) tests are standard for COVID-19 detection but lack precision in assessing pulmonary infection extent.
- Computed Tomography (CT) scans offer a more comprehensive view of lung involvement in COVID-19.
- There is a need for automated tools to analyze CT scans for efficient and accurate COVID-19 diagnosis.
Purpose of the Study:
- To develop a comprehensive, fully-automated system for COVID-19 detection and lesion segmentation from CT scans.
- To leverage deep learning strategies to enhance the decision-making process in COVID-19 diagnosis.
- To improve upon existing methods for analyzing pulmonary infections using medical imaging.
Main Methods:
- A U-shaped neural network architecture with Multiple Convolutional Layers for lung delineation, COVID-19 detection, and lesion segmentation.
- A novel pipeline enabling direct COVID-19 detection and segmentation from CT images.
- A customized loss function designed to balance sensitivity and precision.
Main Results:
- Lung segmentation achieved near 99% sensitivity and a 97% Dice-score.
- The detection network demonstrated 97.1% average accuracy with no false positives across multiple runs.
- Lesion segmentation achieved an average accuracy of approximately 99%.
Conclusions:
- The proposed deep learning system significantly outperforms state-of-the-art methods for COVID-19 lesion segmentation in CT images.
- The algorithm achieved a 38.2% improvement in F1-score compared to the UNet benchmark.
- The high accuracy suggests broad applicability in medical image segmentation for various diseases.
