Weighted Deformable Network for Efficient Segmentation of Lung Tumors in CT

Abstract

The computerized delineation and prognosis of lung cancer is typically based on Computed Tomography (CT) image analysis, whereby the region of interest (ROI) is accurately demarcated and classified. Deep learning in computer vision provides a different perspective to image segmentation. Due to the increasing number of cases of lung cancer and the availability of large volumes of CT scans every day, the need for automated handling becomes imperative. This requires efficient delineation and diagnosis through the design of new techniques for improved accuracy. In this article, we introduce the novel Weighted Deformable U-Net (WDU-Net) for efficient delineation of the tumor region. It incorporates the Deformable Convolution (DC) that can model arbitrary geometric shapes of region of interests. This is enhanced by the Weight Generation (WG) module to suppress unimportant features while highlighting relevant ones. A new Focal Asymmetric Similarity (FAS) loss function helps handle class imbalance. Ablation studies and comparison with state-of-the-art models help establish the effectiveness of WDU-Net with ensemble learning, tested on five publicly available lung cancer datasets. Best results were obtained on the LIDC-IDRI lung tumor test dataset, with an average Dice score of 0.9137, the Hausdorff Distance 95% (HD95) of 5.3852, and Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) of 0.9449.