PulmonU-Net: a semantic lung disease segmentation model leveraging the benefit of multiscale feature concatenation and leaky ReLU

Abstract

Pulmonary diseases impact lung functionality and can cause health complications. X-ray imaging is an initial diagnostic approach for evaluating lung conditions. Manual segmentation of lung infections from X-rays is time-consuming and subjective. Automated segmentation has gained interest to reduce clinician workload. Semantic segmentation involves labelling individual pixels in X-rays to highlight infected regions. This article presents PulmonU-Net, an innovative semantic segmentation model using PulmonNet modules as the base network to highlight infected areas in chest X-rays. PulmonNet modules leverage global and local chest X-ray characteristics to create intricate feature maps. Incorporating leaky ReLU activation enables uninterrupted neuron functioning during learning. By adding PulmonNet modules in the encoder's deeper layers, the model addresses vanishing gradients and improves dice similarity coefficient to 94.25%. Real-time testing and prediction visualization demonstrate PulmonU-Net's effectiveness for automated lung infection segmentation from chest X-rays.