Lung Nodule Segmentation and Classification Using Conv-Unet Based Deep Learning

Abstract

This paper proposes a Convolutional U-Net architecture, a variation of the standard U-Net architecture for the segmentation of lung nodules and classification using Deep learning on Computerized Tomography (CT) scans. The Primary steps employed are Preprocessing, Segmentation and Classification of nodules. In the preprocessing step, the lung region is segmented using techniques such as normalization, median filtering, Kmeans clustering, morphological and thresholding operations to extract lung Region of Interest (ROI) and nodule masks. The Conv-Unet design adds more convolutional layers to the standard U -Net architecture to help capture complicated patterns and boundaries of lung nodules for more accurate segmentation. Categorization of the segmented lung nodules is done using a CNN network on the LIDC-IDRI, and LUNA16 dataset. Overall, this model achieves a dice score of 62% and classification accuracy of 82% displaying appropriate performance in comparison with other variations of the U-Net architecture.