Omnidirectional 2.5D representation for COVID-19 diagnosis using chest CTs.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Coronavirus Disease 2019 (COVID-19) has caused a global health crisis, necessitating advanced diagnostic tools.
• Computerized tomography (CT) scans are crucial for identifying pulmonary infections, but current deep learning methods have limitations.
• Existing 2D deep learning models require extensive per-slice annotations, while 3D models are computationally intensive.

Purpose of the Study:

• To introduce a novel omnidirectional 2.5D representation for volumetric chest CT scans.
• To enable the use of efficient 2D deep learning architectures for COVID-19 detection.
• To reduce annotation requirements to volume-level only.

Main Methods:

• A siamese feature extraction backbone processes each lung independently.
• Features are combined in a classification head utilizing Squeeze-and-Excite strategies and Class Activation Maps.
• The model was trained and validated on public and in-house datasets.

Main Results:

• The proposed 2.5D method achieved prediction quality comparable or superior to state-of-the-art techniques.
• The model accurately differentiated COVID-19 pneumonia from other pneumonias and healthy lung cases.
• The approach demonstrated efficiency by requiring only volume-level annotations.

Conclusions:

• The omnidirectional 2.5D CT representation offers an effective and efficient deep learning approach for COVID-19 diagnosis.
• This method overcomes the limitations of existing 2D and 3D deep learning models.
• The findings support the clinical utility of this novel AI-driven diagnostic tool for pulmonary infections.