Lung cancer remains the leading cause of cancer death worldwide, with approximately 235,000 new cases and 125,000 deaths projected in the U.S. in 2024. Early detection through CT scanning is one of the most powerful tools available, but the volume and complexity of these images overwhelms radiologists. Studies show that underreading errors occur in 42% of cases and cognitive errors contribute to 9% of diagnostic mistakes.
Deep learning models offer a way to automate the detection and delineation of tumors in CT images with radiologist-level accuracy. Unlike manual measurements, AI-driven segmentation can precisely outline tumor boundaries, enabling accurate volumetric analysis and consistent application of RECIST (Response Evaluation Criteria in Solid Tumors) standards. This reduces variability and speeds up the diagnostic workflow.
Despite growing interest in 3D segmentation methods, 2D approaches remain clinically valuable and computationally practical. Clinicians routinely interpret CT images slice by slice in 2D, and 2D deep learning models can leverage large bodies of research from natural image processing. Studies show that multiple 2D models can match or exceed 3D convolution models in segmentation performance while training faster.
This systematic review specifically examines 2D deep learning segmentation networks applied to lung cancer CT scans, covering 124 studies published between 2020 and 2024. It identifies which architectures dominate the field, what datasets are most commonly used, and what critical gaps in methodology need to be addressed before these tools can be widely deployed in clinical settings.
The review followed PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analysis) guidelines to ensure a rigorous and reproducible search process. Researchers queried five major databases including PubMed, Science Direct, Google Scholar, IEEE, and the ACM Digital Library, using standardized search terms combining lung, CT, segmentation, and deep learning keywords.
Strict inclusion and exclusion criteria narrowed the initial pool to 124 qualifying studies. Papers had to involve 2D deep learning segmentation of human CT images specifically for lung cancer diagnosis. Studies were excluded if they focused on non-CT imaging, animal models, COVID-19, radiomics, or were non-English-language publications.
Four independent researchers screened titles and abstracts, with a fifth reviewer resolving disagreements. Full texts were examined to confirm that 2D segmentation was a primary focus and that deep learning was applied beyond simple preprocessing. Of the 124 studies, 17 used both 2D and 3D approaches, but only the 2D components were analyzed in this review.
To assess research quality, the team adapted the TRIPOD guidelines for deep learning and applied 22 yes/no questions to every study. These questions probed dataset clarity, train-test splitting practices, class imbalance handling, cross-validation methods, hyperparameter tuning, model stability, generalizability, missing data handling, and expert involvement in annotation.
UNet and its variants were the most widely used architecture, appearing in 40% of all reviewed studies. CNN-based approaches followed at 23%, attention mechanism models at 12%, and ResNet at 9%. Transformer models, though emerging, appeared in only 2-3% of studies, suggesting a gradual shift toward more advanced architectures capable of capturing long-range spatial relationships.
Supervised learning dominated the field overwhelmingly, used in 90% of all 124 studies. Semi-supervised and unsupervised learning each appeared in only 9 studies, reflecting the field's dependence on fully labeled training data. This is a significant limitation given how labor-intensive expert annotation of medical images is.
The LIDC-IDRI dataset was used in 51% of all studies, making it far more common than any other dataset. This dataset contains 1,018 thoracic CT cases annotated by four experienced radiologists across two independent review phases. The LUNA16 dataset, derived from LIDC-IDRI, was the second most common, containing 888 annotated CT scans focused on nodules 3 mm or larger.
Dice Similarity Coefficient was the most commonly reported performance metric, followed by accuracy, sensitivity, and Intersection over Union. Dice Similarity Loss, Binary Cross-Entropy Loss, and Cross-Entropy Loss were the three most frequently used training objectives. The dominance of these standard metrics reflects the field's maturity but also limits comparisons across studies that report different subsets of metrics.
The best-performing models on the LIDC-IDRI dataset achieved remarkable accuracy, with some reaching Dice Similarity Coefficients of 0.98 to 0.9999. An advanced UNet with Weighted Binary Cross-Entropy Loss achieved a DSC of 0.98 and sensitivity of 99.41%. These results demonstrate that when properly designed and trained, deep learning models can detect lung nodules with extraordinary precision.
Attention-enhanced UNet variants consistently improved performance on difficult segmentation tasks like small or irregularly shaped nodules. AWE-UNet achieved a DSC of 0.91 and RAD-UNet achieved an F1-score of 0.94, both excelling at small nodule detection. Transformer-enhanced models like TransUNet reached a DSC of 0.887, showing promise for capturing contextual information across larger image regions.
On the LUNA16 dataset, models built on UNet with ResNet backbones showed particularly strong results. Preprocessing-UNet-ResNet50 achieved a DSC of 0.99, and R2UNet with densely connected layers reached a DSC of 0.98. These hybrid architectures combining the spatial precision of UNet with the deep feature extraction of ResNet appear especially well-suited for pulmonary nodule segmentation.
Semi-supervised approaches showed competitive performance while reducing dependence on fully labeled data. The Box2Pseudo framework, which combines a small fully labeled dataset with a larger weakly labeled one, achieved an average DSC of 0.87 on LIDC-IDRI, matching fully supervised methods. This approach is particularly relevant given the high cost of expert medical image annotation.
The most pervasive gap identified across reviewed studies was failure to address class imbalance, which affected 67% of all papers. In lung cancer CT images, cancerous regions are typically tiny compared to the surrounding healthy tissue. Without strategies to account for this imbalance, models tend to over-predict the majority (non-cancerous) class, leading to artificially high accuracy but poor detection of actual tumors.
Cross-validation was underused, appearing in only 21% of studies. Without proper cross-validation, it is difficult to assess whether a model's reported performance will generalize to new patients or imaging centers. This gap is particularly concerning given that many studies use relatively small datasets where a single train-test split can produce misleading results.
Model stability testing across different random seeds was performed in only 3% of studies, the lowest rate of any quality criterion examined. This means that the reported performance figures for the vast majority of models may reflect lucky initialization rather than reliable, reproducible performance. Without stability testing, it is impossible to know whether a model would perform consistently in practice.
Missing data was not addressed in 88% of studies, and generalizability was discussed in only 34%. Real-world clinical datasets frequently contain incomplete records, corrupted scans, or missing follow-up data. Models that fail to account for this will likely degrade when deployed outside the controlled research environment. Similarly, most models were evaluated only on the same dataset they were trained on, leaving their real-world applicability unknown.
China dominated global research output, contributing 52 publications representing 31% of the total reviewed output. Together, China and India accounted for more than 50% of all publications in this dataset, reflecting the substantial investment in medical AI research in Asia. This geographic concentration raises questions about whether the models developed on local datasets will generalize to patient populations in other regions.
IEEE was the primary publisher for this research area, accounting for approximately 54% of all publications. Elsevier followed with around 13% and MDPI with roughly 8%. This distribution reflects the technical engineering roots of deep learning research in medical imaging, though it also means many studies appear in conference proceedings that may not undergo the same rigorous peer review as journal publications.
Research output in this field peaked in 2021 before declining through 2024, possibly reflecting a maturing landscape where the most straightforward approaches have been explored. Future growth is likely to come from more advanced architectures, multi-institutional collaborations, and regulatory pathways toward clinical integration rather than incremental improvements to established models.
The concentration of studies on a small number of public datasets creates an artificial benchmark environment that may not reflect clinical reality. LIDC-IDRI and LUNA16 together dominate the training landscape, meaning most models have been optimized and evaluated against the same data. Performance on these standard benchmarks does not guarantee equivalent performance on diverse real-world CT scans from different scanners, institutions, and patient demographics.
The review concludes that UNet and its variants remain the most effective architectures for 2D lung cancer CT segmentation, but they should be integrated into combined 2D/3D modeling workflows. Pure 2D models excel at processing individual slices efficiently, while 3D context helps with capturing volumetric tumor characteristics. A hybrid approach leverages the strengths of both.
Larger and more diverse datasets are urgently needed to train models that can generalize across different patient populations, scanner types, and imaging protocols. The current over-reliance on LIDC-IDRI and LUNA16 limits generalizability and may create models that perform well on benchmarks but fail in real clinical environments. Multi-center and international dataset initiatives are essential next steps.
Expansion of semi-supervised and unsupervised learning methods should be a research priority. Fully labeled medical imaging datasets are expensive and time-consuming to create. Semi-supervised approaches that can learn from partially labeled data, or unsupervised methods that identify structure without labels, would dramatically lower the barrier to training effective models across more diverse clinical contexts.
Future research must prioritize rigorous methodology including cross-validation, class imbalance handling, missing data strategies, and external validation. The field has demonstrated that deep learning can achieve excellent segmentation performance in controlled settings. The next challenge is ensuring that these results are reproducible, generalizable, and reliable enough to trust in real clinical decision-making environments.