ABSTRACT
Lung cancer, being one of the most fatal diseases across the globe today, poses a great threat to human beings. An early diagnosis is significant for better treatment analysis, which is very challenging. The treatment in later stages becomes even more difficult. Due to increasing number of cancer cases, the radiologists are overburdened. The lung cancer diagnosis is mostly dependent on the accurate detection and segmentation of the lung tumor regions. To assist the medical experts with a second opinion and to perform the lung tumor segmentation task in the lung computed tomography (CT) scan images, the authors have proposed an approach based on a densely connected convolutional neural network. In this approach, a 3D densely connected convolutional neural network is used in which dense connections are provided between two convolutional layers, which helps to reuse the features across the layers and is also beneficial for solving the vanishing gradient problem. The proposed network consists of an encoder to capture the features in the CT image and a decoder that reconstructs the desired segmentation masks. This proposed approach is evaluated on an online available dataset for lung tumor, non-small-cell lung cancer-Radiomics dataset, and a dice similarity coefficient of 67.34% is achieved. The proposed approach will assist the radiologists to mark the lung cancer regions in a more efficient manner, and it can be utilized in an automatic computer-aided diagnosis system for lung cancer detection.
