Detection and Stage Classification of UNet Segmented Lung Nodules Using CNN

Abstract

India is experiencing a rapid rise in lung cancer cases, with more than 1 lakh new cases forecast in the next five years. 90% of patients who are not discovered early live for five years, however, only 10-15% of patients who are discovered late live for five years. Lung cancer can be discovered early with computed tomography (CT) imaging. To detect and classify lung nodules and their malignancy levels, Computer Tomography (CT) scans were used on lung patients. Since nodules vary widely in size, shape, and texture, detecting them is a difficult task. Moreover, there are other structures in the lungs that are considered to be non-nodules, for example vessels, fibrosis, diffusive diseases, which often resemble nodules in appearance. Therefore, proposed an end-to-end architecture to recognize nodules from three-dimensional CT volume by modified UNet segmentation and CNN based classification on readily available Lung Image Database Consortium – Image Database Resource Initiative (LIDC-IDRI), virtuously self-governing Indian Lung CT Image Database (ILCID) and LNDb Challenge Dataset. According to the results obtained with the proposed method, nodules can be reliably detected and segmented in contrast to the existing lung segmentation, classification and detection algorithms. For nodule segmentation, attained a Dice Similarity Coefficient (DSC) value is about 0.89 and for nodule detection, achieved the Sensitivity of 98.53%.