Lung Image Analysis using Squeeze and Excitation Based Convolutional Networks

We propose a simple yet parameter-efficient mobile-embedded deep learning architecture, which we term the "Squeeze-and-Excitation based 1D convolutional networks" (SECs) for detecting lung diseases from chest X-ray and Ct-Scan images. Our proposed SEC is guided by considering previous architectures such as SqueezeNet, SqueezeNext, and residual 1D convolutional networks (RCNs). This novel convolutional architecture replaces a 1D CNN layer using the Squeeze-and-Excitation block to construct cost-effective lung disease image classification architecture. Here, we attempt to remove the height-axis 1D convolutional neural network (CNN) layer from the RCNs, which helps to reduce computational complexity. We also apply the Squeeze-and-Excitation block after the width axis 1D CNN layer to boost the network’s performance. Combining both yields our parameter-efficient mobile-embedded architecture, a novel building block that one could stack to form Squeeze-and-excitation-based 1D convolutional networks for image classification. We demonstrate the effectiveness of our model on lung disease (COVID-19 and Pneumonia) detection datasets. The extensive experiment shows that our proposed modifications for 44 layers with widening factor 2 achieve 99.68% and 99.45% validation accuracy for 2 and 3-classes datasets, respectively. They consume 2.2M parameters and 18.4M FLOPS. Moreover, our proposed light network shows state-of-the-art performance on the lung segmentation datasets.