Improving Lung CT Analysis through Fuzzy Dilated Convolution Attention

Computed Tomography (CT) is a versatile imaging modality that can detect and evaluate a wide range of diseases and conditions affecting organs in the thoracic cavity. Computer vision methods, particularly, deep learning architectures using attention mechanisms, have been extensively used for analysis of CT images. However, there is still a need for increased accuracy when it comes to pulmonary CT analysis owing to the complex structures visible in the images. In this study, we propose a novel convolution layer, called fuzzy dilated convolution and develop an attention module. We build an encoder-decoder network using the proposed module and train it on two large publicly available lung CT datasets - LUNA16 and LIDC-IDRI. The performance of the proposed network surpasses state-of-the-art segmentation models on both the datasets. The implementation is available at github.com/Swadesh13/Fuzzy-Dilated-Convolution.