Lung Nodule Segmentation for Explainable AI-based Cancer Screening

Abstract

We present a novel approach for segmentation and identification of lung nodules in CT scans, for the purpose of Explainable AI assisted screening. Our segmentation approach combines the U-Net segmentation architecture with a graph-based connected component analysis for false positive nodule identification. CADe systems with high true nodule detection rate and low false positive nodules are desired. We also develop a 3D nodule dataset that can be used to build an explainable classification model for nodule malignancy and biomarkers estimation. We train and evaluate the segmentation model based on its percentage of true nodules identified within the LIDC dataset which contains 1018 CT scans and nodule annotations marked by four board-certified radiologists. We further present results of the segmentation and nodule filtering algorithm and a description of 3D nodule dataset generated.