Integrated Deep Learning and Stochastic Models for Accurate Segmentation of Lung Nodules From Computed Tomography Images: A Novel Framework

• Published in: IEEE access Vol. 11; pp. 99807 - 99821
• Main Authors: Youssef, Bassant E., Alksas, Ahmed, Shalaby, Ahmed, Mahmoud, Ali H., Van Bogaert, Eric, Alghamdi, Norah Saleh, Neubacher, Alyssa, Contractor, Sohail, Ghazal, Mohammed, Elmaghraby, Adel S., El-Baz, Ayman
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: 3D-UNet; Annotations; Computational modeling; Computed tomography; Computed tomography (CT); Datasets; Deep learning; Deformable models; Fields (mathematics); Image databases; Image segmentation; joint stochastic model; Lung; lung nodules; Lungs; Machine learning; Markov-Gibbs random field; Medical imaging; Metric space; Nodules; Random processes; segmentation; Stochastic models; Stochastic processes; Three-dimensional displays; Tomography; Tumors
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3313174

This paper introduces an innovative model for precise extraction of lung nodules from 3D computed tomography (CT) scans. Our approach comprises two essential preprocessing stages aimed at refining search accuracy and nodule segmentation. Initially, we leverage a two-level joint Markov-Gibbs random field (MGRF) model to delineate the lung region, effectively distinguishing lung wall nodules from the chest region with shared visual characteristics. Subsequently, employing a deep learning U-net technique, we pinpoint the region of interest (ROI) housing the lung nodule, minimizing the inclusion of surrounding lung tissues. Further enhancement comes from a 3D U-net, trained with a novel loss function to mitigate under- or over-segmentation issues. The resulting segmentation robustly outlines lung nodules in terms of morphology and volume metrics, validated by Dice coefficient (DCE), absolute volume difference (AVD), <inline-formula> <tex-math notation="LaTeX">95^{th} </tex-math></inline-formula>-percentile Hausdorff distance (HD), sensitivity, and specificity metrics. To assess our approach, we conducted comprehensive experiments. Our evaluation encompasses in vivo data from 50 patients and employs 679 subjects from the publicly available dataset of the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI). The LIDC-IDRI dataset, a seminal resource for computer-aided diagnosis (CAD) in lung nodules, offers annotations enabling tasks like detection, segmentation, classification, and quantification. Our experiments showcase our model's superiority over existing deep learning methods, particularly evident in metrics such as the <inline-formula> <tex-math notation="LaTeX">95^{th} </tex-math></inline-formula>-percentile HD and DCE. While limited demographic information constrains a comprehensive analysis, our approach's robust performance underlines its potential integration into nodule assessment AI systems.