MFL-Net

MFL-Net classifies COVID-19 from chest computed tomography (CT) images using a lightweight multi-scale feature learning convolutional neural network to support timely diagnosis when RT-PCR testing is limited.

Key Features:

• Lightweight architecture: The network uses an extremely lightweight design with 0.78 million trainable parameters, reducing memory and computational requirements compared to ImageNet-pretrained CNNs.
• Multi-Scale Feature Learning (MFL) blocks: MFL blocks integrate multiple convolutional layers with 3x3 filters and residual connections to extract and preserve multi-scale features across network depths.
• Overfitting mitigation: A sequence of MFL blocks limits model capacity growth and reduces overfitting when training on limited CT datasets.
• Enhanced performance: Evaluations on two publicly available COVID-19 CT imaging datasets report superior performance relative to ImageNet-pretrained CNNs and state-of-the-art methods despite the lightweight design.

Scientific Applications:

• COVID-19 diagnosis from chest CT images: Automated classification of COVID-19 presence from chest computed tomography scans for diagnostic support.
• Rapid screening when RT-PCR is limited: Image-based screening to assist diagnosis in contexts with delayed or scarce RT-PCR testing.

Methodology:

The model is constructed as a sequence of multi-scale feature learning (MFL) blocks that combine multiple 3x3 convolutional layers with residual connections, yielding a 0.78 million-parameter CNN that was trained and evaluated on two publicly available COVID-19 CT imaging datasets and compared against ImageNet-pretrained CNNs and state-of-the-art methods.