Lung Segmentation in CT scans with Residual Convolutional and Attention Learning-based U-Net

Abstract

Lung segmentation is considered as prerequisite step in medical image analysis, particularly for the diagnosis formulation and treatment plan of lung diseases. Hence, we are proposing a residual convolutional and attention learning-based U-Net model for precise and proficient lung segmentation in CT scans. The proposed model incorporates a residual convolutional learning block in place of conventional convolutional layer that is utilized in encoder and decoder and an attention mechanism implemented in skip connections of the conventional U-Net architecture, which resulted in augmenting feature representational capability and advancing the discriminative competence of the model. The model is trained and evaluated on a very well-known public dataset named Lung Image Database Consortium (LIDC) dataset and a private dataset taken from a hospital. Experimental outcomes reveal that the presented model accomplishes state-of-the-art performance in terms of Dice Similarity Coefficient as 0.981 for LIDC and 0.987 for private dataset and outperforms several existing methods. The proposed model has the capability to be employed in various clinical applications including lung disease diagnosis and treatment planning and hence, can assist radiologists in enhancing patient survival rate.