Landscape of 2D Deep Learning Segmentation Networks Applied to CT Scan from Lung Cancer Patients: A Systematic Review

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Lung cancer rates necessitate early detection via computed tomography (CT) scans.
• Deep learning (DL) enhances CT scan analysis for improved lung cancer diagnosis and patient outcomes.
• This review focuses on 2D deep learning networks for lung cancer CT segmentation.

Purpose of the Study:

• To review current and prospective applications of 2D deep learning (DL) networks in lung cancer CT segmentation.
• To summarize existing research, identify key concepts, and highlight research gaps in the field.
• To provide insights into the use of DL for improving lung cancer diagnosis and early detection.

Main Methods:

• Systematic literature search following Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines.
• Analysis of peer-reviewed studies published between January 2020 and December 2024.
• Inclusion of 124 studies focusing on data-driven population segmentation using structured data from major scientific databases.

Main Results:

• The LIDC-LIDR dataset was most frequently utilized, with a strong reliance on supervised learning and labeled data.
• UNet and its variants were the predominant models, achieving high Dice Similarity Coefficients (DSC) up to 0.9999.
• Significant gaps identified include class imbalance (67%), underuse of cross-validation (21%), poor model stability evaluation (3%), failure to address missing data (88%), and limited discussion of generalizability (34%).

Conclusions:

• Convolutional Neural Networks, especially UNet, are crucial for lung CT analysis.
• A combined 2D/3D modeling approach is recommended for enhanced lung CT segmentation.
• Future research should focus on larger, diverse datasets and explore semi-supervised/unsupervised learning for automated lung cancer diagnosis.