Multimodal covid network: Multimodal bespoke convolutional neural network architectures for COVID-19 detection from...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Existing AI tools for COVID-19 prediction are limited to single imaging modalities (CXR or CT).
• A need exists for a unified AI approach capable of analyzing diverse chest imaging data for accurate COVID-19 detection.

Purpose of the Study:

• To develop and evaluate multimodal Convolutional Neural Network (CNN) architectures for COVID-19 detection using both Chest X-ray (CXR) and CT scans.
• To identify optimal neural network parameters and hyperparameters for enhanced diagnostic performance.

Main Methods:

• Development of multimodal CNN architectures, including a bespoke MMCOVID-NET, with varying layers (2-7).
• Systematic evaluation of nine experiments focusing on optimizers, learning rates, and epochs.
• Testing 24 distinct models on both small and large datasets to assess performance and generalizability.

Main Results:

• Four MMCOVID-NET models achieved 100% accuracy on a small dataset.
• MMCOVID-NET-III demonstrated superior performance on a larger dataset, reaching 99.75% accuracy and outperforming state-of-the-art methods.
• Model performance was significantly influenced by the selection of parameters and hyperparameters.

Conclusions:

• Multimodal AI architectures can effectively detect COVID-19 from both CXR and CT images.
• MMCOVID-NET-III represents a highly accurate and robust AI tool for COVID-19 diagnosis.
• Careful tuning of parameters and hyperparameters is crucial for optimizing AI model performance in medical image analysis.