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

Background

The increasing rates of lung cancer emphasize the need for early detection through computed tomography (CT) scans, enhanced by deep learning (DL) to improve diagnosis, treatment, and patient survival. This review examines current and prospective applications of 2D- DL networks in lung cancer CT segmentation, summarizing research, highlighting essential concepts and gaps; Methods: Following the Preferred Reporting Items for Systematic Reviews and Meta-Analysis guidelines, a systematic search of peer-reviewed studies from 01/2020 to 12/2024 on data-driven population segmentation using structured data was conducted across databases like Google Scholar, PubMed, Science Direct, IEEE (Institute of Electrical and Electronics Engineers) and ACM (Association for Computing Machinery) library. 124 studies met the inclusion criteria and were analyzed. Results: The LIDC-LIDR dataset was the most frequently used; The finding particularly relies on supervised learning with labeled data. The UNet model and its variants were the most frequently used models in medical image segmentation, achieving Dice Similarity Coefficients (DSC) of up to 0.9999. The reviewed studies primarily exhibit significant gaps in addressing class imbalances (67%), underuse of cross-validation (21%), and poor model stability evaluations (3%). Additionally, 88% failed to address the missing data, and generalizability concerns were only discussed in 34% of cases. Conclusions: The review emphasizes the importance of Convolutional Neural Networks, particularly UNet, in lung CT analysis and advocates for a combined 2D/3D modeling approach. It also highlights the need for larger, diverse datasets and the exploration of semi-supervised and unsupervised learning to enhance automated lung cancer diagnosis and early detection.

Supplementary Information

The online version contains supplementary material available at 10.1007/s10278-025-01458-x.

Introduction

Lung cancer is a major public health concern due to its high mortality rate [1]. In 2024, the American Cancer Society [2] projects 235,000 new lung cancer cases and 125,000 deaths in the U.S. As the leading cause of cancer death, lung cancer underscores the urgent need for advanced research. Continued studies are crucial for developing more effective treatments, improving early detection methods, and ultimately saving lives in this critical area of public health. Despite declining mortality rates, lung cancer remains the leading cause of cancer death in the U.S. A recent study [3] highlights that mortality declines have outpaced incidence declines in men (5.0% vs. 2.6% annually) and women (4.3% vs. 1.1% annually), driven by advances in treatment, earlier diagnoses, and expanded access to care. However, disparities persist, with Native Americans experiencing the slowest decline and states like Mississippi and Kentucky facing mortality rates two to three times higher than western states, underscoring the need for targeted tobacco control, expanded screening, and equitable access to advanced treatments.

Another study [4] using 2015–2019 data found that lung cancer incidence in women aged 35–54 years was equal to or higher than in men in 40 of 51 U.S. states, with significant differences in 20 states (p < 0.05). This trend was not explained by historical smoking prevalence, which was lower or similar in women in 33–34 states, suggesting other contributing factors. Another study on the worldwide lung cancer population [5] projects that in 2035, incidence rates (ASRs) will decrease by 23% in males (35.8 to 27.6 per 100,000) but increase by 2% in females (16.8 to 17.1). Female ASRs will rise sharply in many countries, nearing male rates in Ireland, Norway, the UK, Canada, and the U.S. The highest ASRs are expected in Belarus (49.3, males) and Denmark (36.8, females). In 2035, new cases in 40 countries are projected to rise by 65%, reaching 2.17 million, with China bearing the largest burden, highlighting the urgent need for strengthened global lung cancer control efforts. Early lung cancer detection ultimately contributes to a successful treatment and prevention and may increase the survival rate in humans [6, 7]. Medical imaging provides 2D (dimensional) or 3D views aiding diagnosis, prognosis, and survival prediction [8–14]. While CT (Computed Tomography) scans provide detailed images used to detect cancer, visual interpretations by radiologists can be challenging due to the difficulty in distinguishing malignant and benign tumors.

visual interpretations by radiologists can be challenging due to difficulty distinguishing malignant and benign tumors. [15, 16]. Underreading errors occur in 42% of cases, and satisfaction of search errors in 22%​. Cognitive errors contribute to 9% of diagnostic mistakes​.These findings highlight the need for improved diagnostic processes, including artificial intelligence (AI) tools to reduce these errors [17]. Medical image segmentation methods have the potential to improve diagnostic outcomes by accurately delineating regions of interest (ROI) and reducing ambiguity in clinical images [9, 18, 19]. Computer-assisted tumor segmentation enables accurate, automated Evaluation Criteria in Solid Tumors (RECIST) and volumetric analyses, leading to more precise assessments of tumor response compared to manual measurements [20].

Volumetric analysis provides additional tumor burden and growth pattern information, useful for assessing treatment response and planning [21–24]. CT imaging is widely used for segmentation due to its accessibility and cost-effectiveness. AI, especially deep learning (DL), automates analysis, saving time and reducing variability and errors in manual measurements [11, 23, 25].

DL advancements have transformed medical image segmentation, achieving radiologist-level accuracy with minimal pre-processing. Models like Convolutional Neural Networks (CNNs), Residual Neural Networks (RNNs), and Generative Adversarial Networks (GANs) efficiently analyze raw images, achieving high accuracy and speed, which has led to widespread adoption [26–29], aiding in tumor assessment and treatment evaluation, and enhancing lung cancer screening by improving nodule detection and classification [30, 31]. Recent reviews on DL-based segmentation methods have mainly focused on either the evolution of deep network architectures or techniques specific to certain architectures. For instance, Zhang et al. [32] delved into sophisticated DL techniques for COVID-19 CT image segmentation, highlighting Transformer structures and categorizing segmentation approaches into encoder-decoder, attention-based, multi-scale/pyramid, and Transformer-based models. Qin et al. [33] explored the automated segmentation of retinal arteries, discussing both traditional and DL approaches with a focus on supervised learning, including CNN, GAN, and UNet-based methods, while offering insights into the strengths and weaknesses of these techniques. Han et al. [34] investigated semi-supervised medical image segmentation, organizing studies into pseudo-labeling, consistency regularization, GAN, contrastive learning, and hybrid methods, and tested several on popular datasets. Khajuria et al. [35] examined reinforcement learning in cancer diagnosis and treatment, reviewing various techniques and challenges. Wang et al. [36] conducted a systematic review on the efficacy of DL models, especially UNet, in identifying brain metastases in Magnetic Resonance Imaging (MRI) images, noting their high segmentation accuracy and sensitivity across different MRI systems. This demonstrates a growing diversity in methods and applications for DL in medical image segmentation.

A review [37] explored federated learning methods to enhance global segmentation performance while maintaining privacy, focusing on brain tumor segmentation and a client-based training strategy. Jiang et al. [38] reviewed deep neural network algorithms for segmenting pulmonary nodules, addressing nodule diversity, segmentation clarity, and environmental factors, and analyzed open-source models using LIDC (The Lung Image Database Consortium image collection) and LUNA 16 (Lung Nodule Analysis 2016) datasets. Ciceri et al. [39] concentrated on DL algorithms for fetal brain segmentation, categorizing algorithms by target structures and highlighting research gaps and future directions for managing fetal MR images. Chen et al. [40] surveyed DL approaches for cerebrovascular segmentation since 2015, covering models like sliding window, UNet, CNNs, and Transformers, and discussed their evolution, challenges, and research opportunities, aiming to guide future studies. Wang et al. [41] reviewed transformer-based methods for brain tumor segmentation, dividing them into pure and hybrid transformer types, and discussed their innovations and benefits. Zhang et al. [29] focused on automating pancreas segmentation using DL for CT and MRI scans, aiming to aid medical image analysis and identify future research avenues. Hung et al. [42] analyzed DL methods for segmenting crucial features in carotid artery ultrasound images, vital for clinical diagnosis. Abbasi [43] examined the progress in DL for stroke lesion segmentation in MRI and CT scans, comparing the performance and noting the strengths and limitations of each modality. Zhang et al. [44] evaluated the effectiveness of DL in segmenting teeth, mandibular canals, and alveolar bones in dental CBCT images, highlighting issues such as dataset inconsistency and lack of method transparency. Xun et al. [45] explored GAN-based architectures for medical image segmentation, highlighting their potential to improve accuracy despite challenges like instability and interpretability issues, which are crucial for wider clinical acceptance. Gu et al. [46] reviewed 2D instance segmentation with DL, focusing on fully-, weakly-, and semi-supervised methods, especially two-stage methods for performance, and provided benchmarks with datasets and metrics for evaluating models.. Instance segmentation is a computer vision task that involves identifying and separating individual objects within an image, including detecting the boundaries of each object and assigning a unique label to each object instance, resulting in a pixel-wise segmentation map of the image [47]. Marques et al. [48] conducted a systematic review on automating Optic Nerve Head segmentation in Optical Coherence Tomography, noting inconsistencies and the need for standardized methodologies for clinical trust and integration. Kaur et al. [49] traced the evolution of medical image segmentation from traditional methods to DL, pointing out the trend towards two-step DL models for multi-organ segmentation and calling for more research on less-studied organs.

The above-mentioned studies show that much of the research in medical image segmentation, especially for CT, centers on 3D techniques. However, the importance of 2D segmentation, particularly in CT imaging, is often overshadowed by the growing interest in 3D methods, despite the critical role that 2D segmentation plays in lung cancer diagnosis. Thanoon et al. [50] and Carvalho et al. [51] emphasize that 2D segmentation offers significant clinical benefits by allowing the processing of individual CT slices, which clinicians use to identify disease patterns with high precision. 2D segmentation models are effective for CT image analysis, offering adaptability, cost-efficiency, and diagnostic accuracy [52]. Despite additional studies on 3D techniques, research [30, 53] has shown 2D segmentation techniques are effective for lung tissue analysis, especially as significant advanced research in 2D-image DL techniques (e.g. natural sceneries) can be utilized and fine-tuned for medical imaging research. The results of the research indicated that employing multiple 2D models lead to quicker training speeds and enabled matched or excelling segmentation performances for 3D convolution models.

These systematic reviews distinguish itself by focusing specifically on the landscape of 2D DL segmentation networks applied to CT scans from lung cancer patients. Unlike previous reviews that may generalize across various DL techniques, we provide a detailed analysis of 2D segmentation networks, their architectures, and their specific applications in lung cancer imaging. Additionally, we highlight the strengths and limitations of existing models. This targeted approach allows us to identify gaps in the current literature and propose future research directions that are specifically relevant to 2D segmentation methods in lung cancer diagnosis. Addressing a gap in studies on 2D segmentation for lung cancer imaging, it explores recent 2D auto-segmentation techniques in DL architectures for CT scans, highlighting developments, advantages, and challenges from 2020 to 2024. Key questions include identifying DL methods for lung cancer segmentation, optimizing 2D segmentation with loss functions, assessing recent advances, and evaluating the benefits and limitations of 2D segmentation in lung cancer analysis. Through a systematic review of recent literature from 2020 to 2024, we focus on answering the following key research questions: (i) what are the most commonly used 2D segmentation models for lung cancer detection in CT images, and how do they perform in terms of accuracy and efficiency? (ii) What are the current gaps in research, and what are the prospects for improving 2D segmentation in lung cancer diagnosis? (iii) What are the advantages and limitations of 2D segmentation methods in lung cancer CT analysis? (iv) What are the strengths and weaknesses of commonly used 2D segmentation architectures, such as UNet and its variants, in segmenting lung cancer lesions from CT images?

Materials and Methods

This article undertakes a systematic review guided by the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) framework. The criteria for study selection and the strategies employed to gather these studies are defined and will be elaborated upon in the sections that follow.

Eligibility Criteria for Study Inclusion

• Eligibility Criteria for Study Selection: For this systematic review, we set clear eligibility criteria to select relevant studies. This review included original research papers and conference proceedings on 2D DL segmentation methods in CT scans for lung cancer diagnosis. Filters were applied to English-language articles published between 2020 and the end of 2024.

• Inclusion Criteria:The criteria for inclusion were:"Segmentation" in the title or abstract with at least two related search terms;Studies involving human subjects;CT images only; andA focus on 2D DL segmentation.

• Exclusion Criteria:Rigorous exclusion criteria filtered out non-CT imaging studies, preclinical/animal studies, testing-only datasets, case reports, reviews, and articles related to radiomics, COVID, or in non-English languages.

• Final Study Selection:Researchers carefully reviewed all articles for relevance, including only studies with 2D DL models for lung cancer segmentation in the final analysis. This stringent selection resulted in 124 studies for detailed review, contributing valuable insights into DL-driven 2D segmentation techniques for accurate, automated lung tumor segmentation in CT imaging.

Search Strategy for Information Source

To locate pertinent studies for a systematic review, an extensive search was carried out from January 1, 2020, to December 31, 2024. A variety of platforms, such as PubMed (pubmed.ncbi.nlm.nih.gov), Science Direct (sciencedirect.com), and Google Scholar (scholar.google.com), IEEE (Institute of Electrical and Electronics Engineers) (https://ieeexplore.ieee.org), and ACM (Association for Computing Machinery) Digital Library (https://dl.acm.org) were employed to access high-quality academic materials including articles from scientific journals and proceedings from conferences.

Study Selection and Data Extraction Process

The study selection process for this systematic review is outlined in the PRISMA diagram shown in Fig. 1 demonstrating the structured and methodical approach used. PubMed yielded 779 papers when searched with the keywords: "segmentation" AND "Lung" AND "CT" OR "computed tomography" OR "deep learning" NOT " COVID ", as shown in Table 1. Science Direct returned 3304 results using the search terms "CT" AND "lung OR cancer OR tumor" NOT "COVID" OR "deep learning." In the ACM library search,94 articles were found based on the search title: segmentation and all: (CT or lung or computed tomography) AND (E-publication Date: (01/01/2020 to 12/31/2024). IEEE produced 759 papers using the same search terms. Finally, Google Scholar returned 41,700 results for the search terms "CT" + "segmentation" + "lung". A more in-depth search of titles within Google Scholar identified 1006 papers that included all of these keywords in the title: "Lung CT segmentation OR cancer OR tumor OR deep learning. To focus on segmentation techniques, "segmentation" was required in article titles. An initial title search and abstract review included the term "lung" to capture studies on lung-related topics like nodules and tumors, while excluding terms like "infection," "COVID," and "organ." Full texts were reviewed to confirm the use of deep learning models beyond pre-processing, and model descriptions were checked to ensure a focus on 2D segmentation for lung cancer CT diagnosis. Four independent researchers screened articles by titles and abstracts, using EndNote to organize references and remove duplicates, with a fifth reviewer resolving any discrepancies for an unbiased selection, we included studies that incorporated both 2D and 3D segmentation. Out of the 124 studies reviewed, 17 included both 2D and 3D models, with our primary focus on analyzing the 2D models. Only original research on 2D segmentation in lung cancer CT imaging meeting the inclusion criteria proceeded to full review.

Fig. 1.

Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) flowchart illustrating the selection process of relevant articles from prominent scientific databases. The symbol 'n' represents the number of articles at each stage of the process

Table 1.

Study Selection Process for Systematic Review on Lung CT Segmentation

Search Database	Search Terms Used (2020–2025)
PubMed	"segmentation" AND "Lung" AND "CT" OR "computed tomography" OR "deep learning" NOT "COVID" AND E-publication Date: (01/01/2020 to 01/01/2025)
Science Direct	"CT" OR "computed tomography" AND "lung OR cancer OR tumor" NOT "COVID" OR "deep learning" AND E-publication Date: (01/01/2020 to 01/01/2025)
ACM Library	segmentation AND all: (CT OR "computed tomography" OR lung) AND (E-publication Date: (01/01/2020 to 01/01/2025))
IEEE	"CT" OR "computed tomography" AND "lung OR cancer OR tumor" NOT "COVID" OR "deep learning" AND E-publication Date: (01/01/2020 to 01/01/2025)
Google Scholar	"CT" OR "computed tomography" + "segmentation" + "lung" AND E-publication Date: (01/01/2020 to 01/01/2025)
Google Scholar—Title Search	"Lung CT OR computed tomography segmentation OR cancer OR tumor OR deep learning" AND E-publication Date: (01/01/2020 to 01/01/2025)
Article Screening	Title and Abstract Review, Full Text Review AND E-publication Date: (01/01/2020 to 01/01/2025) Titles were screened to ensure relevance to lung CT segmentation using deep learning; full texts were reviewed to confirm the use of 2D segmentation techniques for diagnosis

Comprehensive Methodology for Evaluating Research Biases

We conducted a comprehensive review of studies, the most widely used in 2D segmentation of lung cancer in CT images, focusing on identifying research biases. Adapting the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines specifically for DL, we developed targeted questions to rigorously evaluate each study’s strengths, weaknesses, and potential biases.

Question Design for Bias Analysis

To systematically assess critical bias aspects of model development in DL-based studies, we designed Yes/No questions covering key areas where bias might happen. We applied these questions to 124 studies, using the responses to evaluate each study systematically. This helped us identify trends such as gaps in socioeconomic considerations, variability in cross-validation methods, and limited discussions on generalizability. Our findings highlight best practices and gaps, offering insights to enhance UNet model robustness and clinical applicability. These questions addressed key areas, as outlined below and in Table 3. i) Data Sources: Assessed whether datasets were clearly identified as multi-center, single-center, or public, essential for reproducibility and variability understanding. ii) Dataset Splitting: Focused on training, validation, and test set divisions to ensure fair model evaluation. iii) Class Distribution: Checked for class imbalance issues, especially relevant for distinguishing cancerous from non-cancerous areas. iv) Preprocessing Techniques: Evaluated consistency in preprocessing (normalization, cropping, resizing). v) Data Augmentation: Examined the use of augmentation to improve model generalization. vi) Comparison of Models: Reviewed studies comparing models for strengths and weaknesses. vii) Cross-Validation: Ensured proper validation methods, like leave-one-center-out, to prevent overfitting. viii) Hyperparameter Tuning: Reviewed tuning efforts for performance optimization. ix) Model Stability: Assessed stability across random seeds for performance consistency. x) Evaluation Metrics: Evaluated breadth in metrics beyond accuracy and Dice, such as sensitivity, specificity, and AUC. xi) Generalizability: Looked for validations on external datasets to gauge real-world applicability. xii) Case Diversity: Checked for a range of disease severity in cases to improve model versatility. xiii) Socioeconomic and Geographic Factors: Noted studies considering socioeconomic and geographic variability. xiv) Eligibility Criteria: Reviewed clarity in patient selection and treatment descriptions. xv) Handling of Missing Data: Assessed transparency in handling missing data. xvi) Training and Evaluation Datasets: Identified differences between training and evaluation sets to avoid overfitting. xvii) Ground Truth Labeling Consistency: Ensured consistency and minimized annotator bias in labels. xviii) Expert Involvement in Labeling: Checked for multiple annotators to reduce subjectivity. xix) Multi-Modal Data Training: Reviewed use of data from different imaging modalities for robustness. xx) Ablation Study Evaluation: Assessed whether the study performed an ablation study to evaluate the impact of removing certain components or features.xxi) External Dataset Validation: Ensured that the study used external datasets for validation to improve model generalizability.xxii) Statistical Analysis of Metric Differences: Reviewed whether the study conducted statistical.

Table 3.

Comprehensive Analysis of Bias Factors in 2D Segmentation Models for Lung Cancer Across All Studies and UNet Models. This table offers a detailed examination of the principal factors that contribute to bias in 2D segmentation models for lung cancer across all studies and UNet models. It assesses 28 different criteria in various studies, emphasizing elements like dataset transparency, preprocessing standardization, model validation, evaluation metrics, generalizability, and data quality

Question No	Questions	Percent of Yes Answers (%) UNet Models	Percent of Yes Answers (%) All studies
1	Are the data sources clearly identified (e.g., multi-center, single-center, public datasets)?	100	99
2	was the dataset splited (e.g., train, validation, and test sets)?	94	92
3	Are the class distributions (e.g., cancer vs. non-cancer regions) imbalanced?	65	67
4	Were the preprocessing techniques (e.g., normalization, cropping, resizing) standardized across all cases?	100	94
5	Is data augmentation used?	63	60
6	Are different models compared systematically?	90	83
7	Is cross-validation used properly (e.g., leave-one-center-out)?	23	21
8	Have articles referred to hyperparameter tuning methods?	65	68
9	Are models evaluated for stability across different initializations or random seeds?	2	3
10	Number of Evaluation Metrics Reported	-	-
11	At least 1 Evaluation Metric Reported	100	100
12	At least 2 Evaluation Metrics Reported	80	90
13	At least 3 Evaluation Metrics Reported	59	74
14	At least 4 Evaluation Metrics Reported	26	51
15	At least 5 Evaluation Metrics Reported	8	24
16	At least 6 Evaluation Metrics Reported	18	11
17	Does the paper specifically discuss the generalizability of the models to other datasets?	34	53
18	Does the dataset include a wide range of cases, not just severe or advanced cases?	63	71
19	Does model performance vary based on socioeconomic factors or geographic locations?	3	4
20	Are the eligibility criteria and any treatments received by participants described?	5	10
21	Is the handling of missing data described, and its impact on the results discussed?	14	12
22	Are differences between training and evaluation datasets identified?	68	60
23	Are the ground truth labels consistent and free from annotator bias?	0	0
24	Has the ground truth labeling been performed by multiple experts to reduce subjectivity?	65	74
25	Were the CT scans in the dataset obtained using different CT scanners and imaging protocols across various institutions?	70	62
26	Is data used non-public datasets especially for external validation purposes?	17	22
27	Is the study an ablation study ?	89	92
28	Is the study simply bolded or underlined the highest performing metric values without actually doing a statistical analysis of the data to test whether any differences were statistically significant?	25	20

Results

The study also presents findings on the effectiveness of 2D segmentation in CT scans for lung cancer diagnosis, using visual aids for clarity and emphasizing the potential for practical applications and further research.

Distribution of Published DL Articles in 2D Lung Cancer Segmentation and Analysis of DL-Based Segmentation Models Across Different Publishers

Figure 2a features interactive pie charts from 2020 to 2024, showing research publication volumes by country. China's dominance in this area is evident with 52 publications, making up 31% of the total observed research output. China and India have contributed to over 50% of the publications in this dataset over the past few years, highlighting their substantial role in advancing medical imaging research. Figure 2b presents a pie chart illustrating the distribution of research papers on 2D segmentation across the top eight publishers. Each segment represents a specific publisher, with its size indicating the proportion of research papers they have published. IEEE leads with approximately 54% of total publications, highlighting its key role in research dissemination in this field. Elsevier follows with around 13%, MDPI with roughly 8%, Springer with approximately 7%, Science -Direct with roughly 6%. This concentration of publications among a few major publishers underscores their substantial influence in the field.

Fig. 2.

a Distribution of published articles by year and country This bar chart represents the number of research contributions from different countries across the years 2020 to 2024. The stacked bars show the distribution of contributions from various countries, including Australia, Austria, Bangladesh, Brazil, Canada, and others. The line plot overlaid on the chart indicates a trend in the overall contributions, peaking in 2021 and showing a decline through to 2024. b Publication distribution by publishers This pie chart displays the distribution of research publications across various publishers The chart shows that IEEE holds the largest share with ~ 54%, followed by Elsevier at ~ 13%, and other publishers such as MDPI, Springer, Science Direct, ACM Library, Arxiv, Frontiers, and PubMed making up smaller portions of the total distribution

Distribution of 2D Supervised, Semi-Supervised, and Unsupervised Methods in Lung Cancer Segmentation Tasks and Comparison of 2D vs. Combined 2D & 3D Segmentation Techniques in Lung Cancer Research

Figure 3a showcases a bar graph analysis of learning approaches in 2D segmentation studies for lung cancer, revealing a dominant use of supervised learning in 124 instances, accounting for 90% of all methods applied. In contrast, semi-supervised and unsupervised learning methods are minimally used, each found in only 9 studies. This stark differentiation highlights the field's heavy reliance on supervised learning, which prefers models trained on labeled data. The limited use of semi-supervised and unsupervised learning suggests challenges like scarce unlabeled data or the complexity of applying these approaches to medical imagery. Despite the dominance of supervised learning, the findings indicate room for growth in semi-supervised and unsupervised learning methods, potentially broadening data use to enhance lung cancer diagnosis technology. In this review, we included studies that contained both 2D and 3D segmentation. From a total of 124 studies, 17 featured both 2D and 3D models, and we focused on the 2D models for further analysis. Figure 3b from our systematic review illustrates the distribution of 2D versus combined 2D and 3D segmentation approaches in the study of lung cancer. Out of the total, 124 studies employed 2D segmentation methods, which represents 86.0% of the total count. In contrast, 17 studies utilized a combination of 2D and 3D segmentation techniques.

Fig. 3.

a Comparative frequency of supervised, semi-supervised, and unsupervised methods. Bar chart illustrates the distribution of segmentation model types used in lung cancer diagnosis research. Supervised had the highest count of supervised/unsupervised/semi-supervised at 115, followed by semi-supervised, supervised and unsupervised, and unsupervised. Supervised learning had the highest count among supervised, unsupervised, and semi-supervised approaches at 102. This was followed by semi-supervised learning and then unsupervised learning. This emphasizes the dominant reliance on supervised learning approaches in the field. b Proportion of studies utilizing 2D versus combined 2D and 3D segmentation techniques in lung cancer in this research.The pie chart shows the distribution of 2D versus combined 2D and 3D segmentation approaches in lung cancer studies. 2D segmentation methods are used in ~ 86% of the studies, while a combination of 2D and 3D segmentation techniques is employed in ~ 14%. This highlights the prevalent use of 2D approaches in current research

Distribution of DL Model Usages in 2D Lung Cancer Segmentation

Figure 4 illustrates the frequency of top models used in 2D segmentation studies for lung cancer. The flow diagram emphasizes the dominance of the UNet model, utilized in 40% of studies, making it the preferred choice. CNN follows at 23%, with Attention mechanisms at 12% and Residual Network (ResNet) at 9%. Other models, including Fully Convolutional Network (FCN), Semantic Segmentation Network (SegNet), Mask Region-based CNN (Mask-RCNN), Region-Based CNN (RCNN), and Transformer, each have frequencies between 2 and 3%. This data highlights the research community's strong preference for UNet due to its effectiveness in medical image segmentation, particularly for lung cancer. The significant gap between UNet and other models underscores its reliability, while the variety of models reflects ongoing exploration of DL architectures to improve segmentation accuracy and performance. The emerging interest in Transformer models, although less common, suggests a shift towards advanced architectures for complex medical imaging challenges.

Fig. 4.

Distribution of model usage in 2D segmentation research for lung cancer. This image illustrates the frequency distribution of popular deep learning (DL) models used in various tasks, with UNet being the most frequently utilized model at 40%, followed by Convolutional Neural Network (CNN) at 23%, Attention at 12%, and other models such as Residual Network (ResNet), Fully Convolutional Network (FCN), Mask Region-based Convolutional Neural Network (Mask-RCNN), Semantic Segmentation Network (SegNet),Visual Geometry Group (VGG), Region-Based Convolutional Neural Networks (RCNN), and Transformer with lower frequencies. The chart highlights the dominant role of UNet and CNN in the selected domain

In the domain of computer-aided diagnosis and image segmentation for lung cancer detection, this review delves into leading DL models that have revolutionized the field by improving the accuracy, precision, and efficiency of identifying malignant areas in medical images. Notably, the following models stand out as pivotal advancements in 2D segmentation for lung cancer in CT images. Additional descriptions and categorizations of these models, along with related studies on 2D DL segmentation methods, are provided in Supplemental File 1. Tables S1 and S2 in Online Resource Materials provide a detailed summary of the various 2D-DL networks employed for lung cancer segmentation in CT scans, including the segmentation approach utilized, their respective advantages and limitations, as well as the datasets leveraged for training and evaluation. Additionally, Table 2 provides a summary of Supplemental File 1 (Supplemental Table S2) in the Online Resource.

Table 2.

Papers Reviewed for 2D Lung Cancer Segmentation Using deep learning (DL) Models. The table summarizes the key aspects of each study, including the DL model architecture employed, the datasets used for training and evaluation, and the reported performance metrics. Sen: Sensitivity, Specificity: Spe, Accuracy:ACC, JAC: Jaccard, F1-scores: F1-s, Precision = Prec, Recall: Rec, Spearman correlation coefficient: SCC, SVM: Support Vector Machine, interstitial lung diseases: ILDs, Deep Convolutional Neural Network based Segmentation with Two-Pass Contour Refinement: DCNN-TPCR, two-stage Region Proposal Network: RPN, Medical Segmentation Decathlon: MSD, Hausdorff distance(pixels): HD, mean Sensitivity: MSen, mean Precision: MPrec, Fully Convolutional Network: FCN, Binary Cross Entropy: BCE, Fast Multi-Crop Guided Attention: FMGA, Deep Residual Separable Convolutional Neural Network: DRS-CNN, The National Lung Screening Study: NLSS. Peak signal-to-noise ratio: PSNR, Structural similarity index: SMMI, False Positive Rate: FPR, Positive Predictive Value: PPV. The National Lung Screening Study: NLSS. CapsNet: Capsule Network, SC-BLSTM: Spatial-Context Bidirectional Long Short-Term Memory, AC-UNET: Attention Constrained UNet, CSC-Net: Cancer Sensitive Cascaded Network, AEC-Net:Attention and Edge Constraint Network, BPNN: Back Propagation Neural Network MRR:Multi-Resolution Representation, DU-Ne: Dual UNet, ViFUNet:Vision Flash-Based UNet, SMR-UNe:Scale-Aware, Structural Measure: S-M, Enhanced Measure: E-M, Multiattention-Guided Reverse Network, DMC-UNT: Dilated Multi-scale Convolutional UNet, Interstitial Lung Diseases: ILD, Jaccard Similarity:JS, ResDSda_UNet: Residual UNet with Deep Supervision and Dense Connections, SDC:Sorensen-Dice Coefficient Rows with green-colored studies are related to the LUNA16 dataset, while those with yellow-colored studies are related to the LIDC-IDRI dataset. The blue-colored rows correspond to the LUNA 16 and LIDC/IDRI datasets. The maximum values of the evaluation metrics reported in the table. Bias score (Total score of 28) is the sum of "Yes" answers for each study. This Total is provided for each study in Supplemental File 2, with the corresponding details

Usage of Evaluation Metrics in DL-Based 2D Lung Cancer Segmentation Tasks

Figure 5 uses a bar chart visualization to depict the range of evaluation metrics for 2D segmentation models in lung cancer diagnosis. This bar chart represents the frequency of usage of various evaluation metrics in medical imaging research. The most frequently used metric is Dice Similarity, followed by Accuracy, Sensitivity, and Intersection over Union (IoU). Other commonly used metrics include Precision, Specificity, and Jaccard Index, with fewer studies utilizing metrics like Hausdorff Distance, False Positive, and False Negative.

Fig. 5.

Distribution of Evaluation Metric Usage in 2D DL segmentaion in lung cancer for CT imaging This bar chart represents the frequency of usage of various evaluation metrics in medical imaging research. The most frequently used metric is Dice Similarity, followed by Accuracy, Sensitivity, and Intersection over Union. Other commonly used metrics include Precision, Specificity, and Jaccard Index, with fewer studies utilizing metrics like Hausdorff Distance, False Positive, and False Negative

Figure 6 outlines the top six varieties and frequency of loss functions used in developing 2D segmentation models for lung cancer, as identified in this systematic review. This visualization presents the results of applying different loss functions to a 2D segmentation task for lung cancer images, showcasing the frequency of each function's use. The Dice Similarity Loss, Binary Cross-Entropy Loss, and Cross-Entropy Loss occupy the largest portions of the chart, indicating their prominent roles in driving the segmentation models to learn the spatial and structural relationships within the lung cancer images. Other loss functions, such as Focal Loss, Weighted Cross-Entropy, and GAN Loss, are also present, albeit in smaller portions. They contribute by addressing specific challenges like hard-to-segment regions or rare class occurrences. The variation in the use of these loss functions reflects the complexity and multifaceted nature of lung cancer image segmentation, where different loss functions emphasize different aspects of model performance, from accurate tumor detection to minimizing false negatives.

Fig. 6.

The bar chart illustrates the frequency of utilization of different loss functions in medical imaging research. The loss functions are listed on the y-axis, and their corresponding usage frequency is represented by the length of the bars on the x-axis. The most frequently used loss function is the Dice Similarity Coefficient Loss, followed by Binary Cross-Entropy Loss and other types of loss functions such as Cross-Entropy Loss, Focal Loss, and Weighted Cross-Entropy Loss

Distribution of Datasets in DL-Based 2D Lung Cancer Segmentation

This paper section focuses on the usage frequency of top 10 different datasets in 2D lung cancer segmentation research, as shown in Fig. 7. A systematic review illustrated by a bar chart identifies LIDC-LDIR as the top-used dataset, accounted for 51% of Sum of Count of dataset.. The LIDC-IDRI dataset, created through collaboration between seven academic centers and eight medical imaging companies, contains 1,018 cases. Each case includes clinical thoracic CT images and a file documenting a two-phase annotation process by four experienced thoracic radiologists. In the blinded-read phase, radiologists independently identified lesions in three categories: "nodule ≥ 3 mm," "nodule < 3 mm," and "non-nodule ≥ 3 mm [50, 51, 53, 54] The second most used dataset for 2D segmentation after dataset LIDC-IDIR is LUNA16. The LUNA16 dataset is a dataset for lung segmentation. Derived from the LIDC-IDRI dataset, it contains 888 CT scans with annotations from four radiologists, categorizing nodules as solid, part-solid, or non-solid, focusing on nodules that are 3 mm or larger in diameter. It includes annotations confirmed by four experienced radiologists, categorizing lesions into nodules ≥ 3 mm, nodules < 3 mm, and non-nodules.

Fig. 7.

Top 12 frequency distribution of datasets Bar chart displays the distribution of datasets used in lung cancer segmentation studies. The LIDC-IDRI and LUNA16 datasets are the most frequently used. The chart highlights the dominance of specific datasets in lung cancer 2D segmentation research, reflecting their importance and widespread application in the field. The number above each bar represents the number of subjects related to the dataset

Analysis of the Most Commonly Used LIDC-IDRI and LUNA16 and Datasets Across DL Segmentation Models

In lung nodule segmentation using the LIDC-IDRI dataset, various UNet-based models have shown distinct performances. Li Dong et al. [55] proposed an improved U-Net + + model with a ResNeXt encoder and SCSE attention module for lung nodule segmentation in CT images. This approach enhanced feature extraction and achieved a sensitivity of 99.41%, specificity of 99.97%, a false positive rate of 0.02%, and a Dice Similarity Coefficient (DSC) of 0.98, indicating high accuracy in identifying cancerous nodules. The Deep Feature Decoupling Module (DFDM) improves lung nodule segmentation by separating boundary, noise, and texture information from CT images. Integrating DFDM into the UNet architecture enhanced segmentation performance, increasing IoU by 1.51% and Dice score by 1.02% [56]. Kumar et al. [57] propose an ensemble DL approach for lung cancer detection, combining UNet for segmentation and a convolutional neural network for classification. The model achieves a sensitivity of 90.83%, specificity of 94.28%, precision of 93.1%, and accuracy of 91.23%. The Incremental-MRRN and Dense-MRRN models enhance lung tumor detection accuracy while preserving edges, achieving a bit error rate of approximately 0.015%. These models outperform existing algorithms by effectively reducing noise and artifacts in CT images [58]. A Dual Encoding Fusion Network (DEF-Net) for segmenting atypical lung nodules in CT images, achieving an average Dice coefficient of 0.85. DEF-Net outperforms existing state-of-the-art methods, demonstrating its effectiveness in addressing segmentation challenges [59]. Box2Pseudo is a semi-supervised framework for pulmonary nodule segmentation, utilizing a small fully labeled dataset (pixel-level and bounding box annotations) and a larger weakly labeled dataset (bounding box labels only). It comprises three networks: the box-prompt network (BPN), pseudo-refine network (PRN), and main network (MAN), along with a background-filter layer (BFL) to improve pseudo label quality. On the LIDC-IDRI and HX-NODULE datasets, Box2Pseudo achieves an average Dice coefficient of 0.87, outperforming state-of-the-art methods and matching fully supervised performance [60].

The models were utilized for lung nodule segmentation on the LIDC-IDRI dataset, each exhibiting unique strengths and weaknesses. Advanced UNet with Weighted Binary Cross Entropy Loss achieved a high DSC of 0.98 but requires large datasets [61]. The standard UNet scored 0.82 in DSC, though it lacks malignancy classification [62]. More advanced variants like AWE-UNet (DSC: 0.91) [63] and RAD-UNet (mIoU: 0.88, F1: 0.94 [64] improved segmentation, particularly for small nodules. AMST achieved high sensitivity (0.97 for LIDC, 0.94 for IDRI) but may miss some nodules [65]. Dilated Convolutional FCN performed well with a DSC of 0.970 but is limited by 2D kernels [66]. LDDNet showed excellent recall and sensitivity (0.99), though it is time-consuming [67]. MAN-SVM reached a sensitivity of 0.980 and F1-score of 0.98 but requires validation on larger datasets [68]. DB-ResNet reported a DSC of 0.83 but demands significant data and computational power [69]. Lung_PAYNet achieved a DSC of 0.96 using attention mechanisms, while Global-Edge Dual-Path improved segmentation with a DSC of 0.90 [70].

Advanced UNet with Weighted Binary Cross Entropy Loss achieved [61] the highest DSC of 0.98, demonstrating robust segmentation accuracy. AWE-UNet [63] (DSC: 0.90) and RAD-UNet [64] (F1-score: 0.94) also performed well, particularly in attention mechanisms and small nodule segmentation. TransUNet (DSC: 0.887) and Wavelet UNet + + [71] (DSC: 0.94) improved feature extraction using transformer-based architectures and wavelet pooling, respectively. Other models, like Dilated Convolutional FCN (DSC: 0.97) and LDDNet [67] achieved strong performances but are computationally expensive(Recall: 0.97). AMST [65] and Lung_PAYNet [70] offered high accuracy but are constrained by reliance on accurate segmentation and large datasets. Each model brings unique contributions, with some excelling in accuracy, while others focus on efficiency or handling challenging nodules Thakur et al. [72] investigate uncertainty estimation for switch learning based on active learning (TLAL) in clinical image segmentation, utilizing a semi-supervised recurrent convolutional neural network (RCNN) to enhance model accuracy with minimal manual annotation. Their method introduces a novel uncertainty score and demonstrates significant improvements on lung CT datasets, outperforming traditional techniques by over 10% in Dice scores. Various models for lung nodule segmentation using the LIDC-IDRI dataset exhibit unique strengths and limitations. Overall, these models contribute to improved accuracy and efficiency in lung nodule detection, each targeting specific challenges within the field.

Table 2 presents a summarized version of Table S2 (Supplemental File 1) , as shown below. The various studies on 2D segmentation methods for lung cancer, primarily utilized the LIDC-IDRI dataset. Key approaches include Fuzzy C-Means, UNet variants, CNNs, and semi-supervised methods, yielding notable metrics such as high sensitivity (99.41%) and accuracy (up to 98.71%). Many studies emphasize the advantages of enhanced segmentation accuracy and efficiency, along with effective multi-scale feature extraction. However, challenges such as high computational complexity, potential false positives, and limitations in distinguishing benign from malignant nodules are common. Notable papers include those detailing innovative frameworks and DL models, with links provided for further exploration. Overall, these studies contribute significantly to improving lung cancer detection through advanced segmentation techniques.

As shown in Table 2, in lung cancer segmentation using the LUNA16 dataset, a wide range of models based on the UNet architecture and its derivatives have been explored, demonstrating strong segmentation performances. R2UNet [73], with its densely connected layers, reported a DSC of 0.98, reflecting its exceptional accuracy in feature reuse and lung nodule segmentation. Similarly, DC-UNet [74], employing dilated convolutions, achieved a DSC of 0.745, indicating moderate segmentation capabilities but facing challenges with irregular nodule shapes. The basic UNet [75] model has been a foundational architecture, achieving consistently high accuracy metrics, including an impressive F1-score of 0.99 and accuracy of 1.00. Enhancing the UNet structure, Preprocessing-UNet-ResNet50 [76] integrated ResNet50 as a backbone for feature extraction, resulting in superior performance with a DSC of 0.99, making it one of the top-performing models.

On the other hand, AWE-UNet [63], another UNet-based architecture, reported a DSC of 0.90, balancing segmentation accuracy and computational cost. The Improved UNet with Alpha-hull correction [77] yielded a DSC of 0.86, focusing on refining boundary segmentation, though it faced generalization challenges due to the limited dataset size. Combination of UNet and ResNet34 [78] and UNet-ResNet34 [79] further improved performance with residual learning, reporting DSCs above 0.93 for multiple datasets, demonstrating their robustness across various lung pathologies. SW-UNet [80] achieved a DSC of 0.84, offering high accuracy (0.99) but requiring substantial computational resources. U-Det [81] achieved a DSC of 0.83, emphasizing its efficacy in segmenting small and complex nodules, while Modified UNet-based [82] achieved a DSC of 0.88, reflecting strong performance in lobe segmentation. VLSM-Net [83], combining DL with level set refinement, produced a DSC of 0.82 for LUNA16, excelling at detecting blurred boundaries, and UNet# [84] demonstrated high accuracy with a DSC of 0.98 on the Lits17 dataset, highlighting its capability to reduce false positives. Lastly, SquExUNet [85] reported a DSC of 0.80, leveraging cascaded CNNs but struggling with irregular nodule shapes.

In contrast, other models employed for LUNA16, such as Mask R-CNN with K-means [86], focused on unsupervised detection, though its DSC was not reported, and it struggled to differentiate between cancerous and non-cancerous nodules. Faster R-CNN and RetinaNet [87] achieved moderate precision (0.357 and 0.34, respectively) in lung nodule detection, offering limited segmentation accuracy compared to the UNet-based models. The Cascaded Multi-stage UNet [88] performed well with an accuracy of 0.95, though it missed low-contrast nodules. DB-NET [89] offered a high DSC of 0.89, positioning it as a reliable segmentation model. FCN [90] and T-Net [91] demonstrated high accuracy and sensitivity, with T-Net reporting an accuracy of 0.992 and a F1-score of 0.99, though it was limited to 2D images. Finally, MDFN [91] achieved a DSC of 0.89 and strong edge detection, though it required further validation against existing segmentation models. Tang et al.'s SM-RNet [92], with innovative scale-aware multi-attention and reverse erasure, achieves 89.29% Dice on FUSCC and 86.50% on LUNA16, outperforming state-of-the-art model. The Improved UNet model [93], achieved a DSC of 84.48, showcasing enhanced accuracy and processing speed. However, it faced challenges related to its complexity, which could lead to overfitting if not managed with a sufficiently large and diverse dataset. Following this, the ViFUNet (Vision Flash-based UNet), model [94], proposed in 2022, outperformed traditional UNet architectures by reaching a DSC of 86.03. This model benefits from a faster convergence speed and a reduced parameter count, making it efficient for processing large nodule datasets. Nonetheless, it may struggle with very small nodules, which can limit its clinical applicability. Another study in 2024 [95] highlighted a classic UNet model, achieving an impressive Intersection over Union (IoU) of 0.95. This model demonstrated high accuracy in segmenting lung structures, although it also required substantial annotated datasets and computational resources for training.

In summary, UNet-based models, particularly Preprocessing-UNet-ResNet50, R2UNet, and UNet#, demonstrated the highest DSC values, offering robust and accurate lung nodule segmentation on the LUNA16 dataset. Models like SquExUNet and DC-UNet showed moderate performance, while others like DB-NET and MDFN performed well in specific tasks but required further validation for broader application. Non-UNet-based models such as Faster R-CNN and Mask R-CNN delivered moderate results but generally underperformed compared to UNet and its derivatives. A Semantic Segmentation of CT Scans study employed a UNet model, achieving an IoU of 0.95, indicating high accuracy and efficiency in lung structure segmentation, but it also highlighted the need for large annotated datasets for effective training [96].

Beyond the UNet family, other models have contributed to the landscape of lung cancer segmentation. The Deep Neural Networks for Semantic Segmentation study from 2021 employed a CNN-based approach, reporting an IoU of 0.85. While it achieved high segmentation accuracy, it also demanded considerable computational resources, which can be a barrier for implementation in resource-constrained settings [97]. The Detecting Lung Nodules Based on DL model utilized a combination of UNet, CNN, and PCA in 2023, achieving an accuracy of 95.92% for the CNN + PCA approach. This model offered a reduced computational cost and a simpler architecture; however, it was criticized for potentially losing spatial context in the segmentation process [98] (Table 2).

Bias Analysis in Lung Cancer Segmentation Studies Using CT Images: A Focus on UNet Models

Our analysis of 124 studies on lung cancer segmentation, assessed using tailored TRIPOD questions, revealed critical insights summarized in Table 3. Regarding Data Handling, while all studies (100%) report data sources (Q1) and 92% detail dataset splitting (Q2), only 67% address class imbalance between cancerous and non-cancerous regions (Q3). In Preprocessing and Augmentation, standardized preprocessing is universally applied (94%, Q4), but only 60% enhance dataset diversity through data augmentation (Q5). For Model Optimization, models’ comparisons are present in 83% (Q6), but only 21% use cross-validation, risking overfitting (Q7). Hyperparameter tuning is mentioned in 68%(Q8), while model stability across initializations is rarely assessed (3%, Q9). Evaluation Metrics reveal 100% report at least one metric (Q10–Q11), yet only 11% use six metrics(Q16), reflecting limited evaluation depth. On Generalizability and Diversity, just 53% address external dataset generalizability (Q17), 71% include diverse cancer stages (Q18), and only 4% consider socioeconomic or geographic factors (Q19). Finally, under Annotation and Protocols, 74% reduce subjectivity with multiple annotators (Q24), and 76% incorporate varied CT scanners and protocols (Q25) for robustness. However, key gaps persist, with only 10% discussing eligibility and treatments (Q20), 12% addressing missing data (Q21), 60% acknowledging training-evaluation dataset differences (Q22), and none ensuring ground truth consistency (Q23)0.74% use multiple annotators (Q24) and 62% usedata from Varied CT scanners and protocols(Q25). Only 17% of the reviewed studies used non-public datasets for external validation, indicating a need for broader validation to ensure generalizability (Q26). A significant 89% of reviewed studies performed ablation studies, highlighting a commitment to understanding the impact of model components (Q27). Additionally, 25% of studies highlighted top metrics without statistical analysis, underscoring a need for more rigorous statistical validation in reported results (Q28).

We have added a bias analysis specifically for UNet models, along with a question analysis, as these are the most frequently reported models in 2D segmentation of lung cancer CT imaging. Please refer to Table 3 for more details.

For all studies, similar strengths and weaknesses were observed. Data source identification remained high, and dataset splitting was also widely used. However, 67% of studies reported class imbalances, and 96% standardized preprocessing techniques, demonstrating strong methodological consistency. Cross-validation was underused (21%), and model stability was poorly evaluated (3%). While most studies reported at least one evaluation metric, a significant portion did not report enough metrics for thorough performance assessment. Generalizability was discussed in only 34% of studies, and socioeconomic factors or geographic variations were largely ignored. The handling of missing data was not addressed in most studies (12%), and external validation using non-public datasets was limited to 22%.

For UNet models, the studies demonstrate strong transparency in data source identification and dataset splitting, ensuring reproducibility and model validation. Preprocessing techniques were standardized in all studies, which is a significant strength. However, class imbalance was reported in 65% of studies, highlighting a common challenge. Data augmentation was used in 63% of studies, but its implementation could be expanded. Notably, only 23% of studies used cross-validation, and only 2% evaluated model stability across different initializations. While the majority of studies reported at least one evaluation metric, there was a gap in reporting a broader range, with only 18% reporting six or more metrics. Generalizability to other datasets was addressed in 53% of studies, but there was little focus on the impact of socioeconomic or geographic factors on model performance. Furthermore, missing data was handled in just 14% of studies, and only 65% ensured ground truth labels were reviewed by multiple experts to reduce bias. The bias score (total score of 28) is the sum of "Yes" answers for each study. This total is provided for each study in Supplemental File 2, with the corresponding details. Additionally, the sum of "Yes" answers for each study is included in Table 2.

Discussion

Lung cancer is the leading cause of cancer-related deaths worldwide with a five-year survival rate remaining alarmingly low due to late-stage diagnoses [1, 2]. Early CT screening has been proved to reduce mortality rates by enabling early detection and then intervention. Recent advancements in AI, including radiomics, machine learning (ML), and DL, have significantly transformed imaging analysis, particularly in evaluating lung nodules. Automated segmentation has become a critical component of quantitative analysis, streamlining the detection, diagnosis, and monitoring of lung nodules by reducing time, labor, and variability. In medical imaging, 2D segmentation plays a vital role in accurately extracting ROIs. For instance, a recent study [182] indicated that 2D slice segmentation remains clinically important as it captures micro -environmental heterogeneity that may be overlooked in 3D volumetric analysis. Research has demonstrated that peritumoral radiomic features extracted from 2D CT slices can independently predict treatment response in lung cancer patients, showing comparable prognostic value to 3D analyses. Additionally, 2D features enable a more granular, slice-by-slice assessment of lesion characteristics, providing statistical insights into texture variations that a single 3D value may miss. This approach is particularly valuable when full 3D segmentation is impractical in clinical workflows and supports future optimization studies using dominant lung lesions for prediction.

The rapid evolution of AI technologies, such as radiomics and DL algorithms, has enhanced quantitative imaging analysis for lung cancer, particularly in lung nodule imaging. Segmentation serves as the cornerstone of workflows ranging from nodule detection to diagnosis and ongoing monitoring. Automated nodule segmentation has garnered significant attention due to its ability to address challenges associated with time-intensive workflows, laborious manual effort, and susceptibility to human error and variability.

Automated 2D segmentation offers several distinct advantages. It aligns well with most medical imaging databases, which predominantly store 2D images. Unlike 3D models, it eliminates the need for resizing, ensuring seamless analysis. Additionally, it reduces processing time, enabling quicker review of CT slices, and offers flexibility for re-training across diverse 2D image datasets, including X-rays. These benefits make 2D segmentation a practical and scalable solution for clinical workflows [182, 183].

While 3D segmentation methods are increasingly favored for their ability to capture volumetric tumor characteristics and inter-slice relationships, 2D segmentation remains highly relevant and indispensable in medical imaging due to its accessibility, computational efficiency, and alignment with slice-based diagnostic workflows. It is particularly valuable and practical in resource-limited settings where 3D data acquisition or processing may be impractical, allowing for rapid analysis, easy adaptation across datasets, and serving as a foundation for advanced techniques like semi-supervised learning. One of its key advantages is feasibility with limited data, as annotated 3D datasets are costly and time-intensive to acquire, whereas 2D methods rely on slice-by-slice annotations, making them more accessible. Training 2D models on individual slices extracted from 3D volumes increases the number of training samples, and by integrating these slices, we can reconstruct volumetric segmentation, addressing data scarcity and enhancing generalizability. Although 3D CNNs can capture intricate volumetric features, they require substantial computational resources and extended training times. By contrast, analyzing volumetric images slice-by-slice using 2D CNNs, combined with data augmentation and tri-planar views, helps preserve essential 3D characteristics. This approach strikes a balance between efficiency and effectiveness, underscoring the importance of 2D segmentation in advancing lung cancer imaging and treatment workflows. Additionally, 2D segmentation provides a steppingstone for semi-supervised and transfer learning strategies, bridging the gap between 2D and fully 3D segmentation tasks. Its continued clinical relevance in assisting radiologists with tumor identification and precise image analysis highlights its enduring importance alongside emerging 3D methodologies [93, 183].

Pang et al. [184] emphasized the effectiveness of 2D segmentation in addressing critical challenges such as detecting small ROIs and capturing subtle anatomical details, as demonstrated in liver tumor and retinal vessel segmentation tasks. Furthermore, 2D methods are more practical in scenarios where 3D data is unnecessary or difficult to obtain, requiring less computational power, simpler implementation, and benefiting from readily available pretrained models to accelerate development. This flexibility makes 2D segmentation an essential tool in many computer vision applications, particularly when balancing efficiency and effectiveness is a priority.

X-ray radiographs, among the most widely used imaging modalities in ML and DL studies, further demonstrate the importance of 2D segmentation. Radiographs are common in both developed and developing countries and serve as the basis for diagnosing numerous diseases. Publicly available datasets, such as the CheXpert and NIH Chest X-Ray datasets, have driven significant advancements in 2D segmentation techniques. Despite the initial limitations of 2D representation, such as reduced soft tissue contrast and the loss of volumetric context, 2D segmentation has laid the groundwork for more sophisticated, non-invasive imaging modalities and remains pivotal in advancing research and clinical practice in lung cancer diagnosis [185].

Additionally, while 2D segmentation techniques can extend to continuous video data by treating it as a series of 2D frames, this approach sacrifices the ability to leverage temporal continuity inherent in video. This trade-off highlights the need to balance computational efficiency with the incorporation of advanced temporal modeling for applications requiring sequential data processing [186].

Despite the shift towards 3D methods, the enduring value of 2D segmentation lies in its ability to address specific clinical needs effectively and efficiently. By refining 2D segmentation methods to mitigate limitations such as the lack of inter-slice correlations, their clinical relevance can be further enhanced, complementing 3D segmentation approaches while maintaining their unique advantages. This study underscores the foundational role of 2D segmentation in medical imaging, offering a bridge between accessible, efficient workflows and advanced diagnostic capabilities.

This study presented a review of 2D lung segmentation in CT images, answering the research question: “What are the 2D segmentation methods for the lung cancer, using CT images”? This analysis examined the latest developments in 2D segmentation techniques for lung cancer CT scans, paying special attention to the growing influence of DL applications in this area. Findings highlight a strong shift towards DL, particularly supervised learning, driven by public datasets like LIDC-IDRI and LUNA16 and, which foster innovation. The analysis reveals satisfactory accuracy across 2D segmentation techniques, with UNet models often outperforming others due to its adaptability to complex shapes. The DSC is a popular metric for 2D segmentation, measuring overlap between the segmented image and ground truth to assess accuracy. It balances false positives and negatives, providing a robust evaluation. However, DSC can be sensitive to small variations, impacting robustness in cases requiring precise delineation [187, 188].

Although DL shows promising results, evaluating and comparing 2D segmentation approaches is challenging due to: (i) Differences in Study Methodologies: Variations in methods, sample sizes, inclusion criteria, and benchmarks lead to inconsistencies, complicating direct comparisons of performance metrics; and (ii) Absence of Standardized Metrics. The lack of uniform evaluation metrics and reporting protocols across studies further hinders the conclusive identification of the best approach. In lung cancer detection, performance metrics such as DSC, sensitivity, specificity, and accuracy are critical for evaluating diagnostic models. The DSC measures the overlap between predicted and actual tumor regions, with higher values indicating better segmentation accuracy. Sensitivity reflects the model's ability to correctly identify true positives, which is essential for early detection of malignant lesions, potentially improving treatment outcomes. Specificity assesses the model's capacity to correctly identify true negatives, thereby minimizing false positives and avoiding unnecessary diagnostic procedures. While accuracy provides an overall measure of correct predictions, it can be misleading in imbalanced datasets, as it may not adequately reflect true model performance. Müller et al. [187] highlight the importance of achieving an optimal balance between sensitivity and specificity to enhance patient management and ensure reliable diagnostic applications in clinical practice. However, as Liu et al. [189] point out, task-agnostic metrics like DSC, Jaccard Similarity Coefficient (JSC), and Hausdorff Distance (HD), though commonly used, do not always correlate with clinically relevant outcomes. Their study underscores the need for task-based evaluation of segmentation algorithms, particularly in assessing quantitative measures such as metabolic tumor volume (MTV) and total lesion glycolysis (TLG). They demonstrated that algorithms performing well on DSC may exhibit significant biases in estimating clinically important metrics. For example, deeper U-Net models improved MTV and TLG estimation accuracy despite showing similar performance on task-agnostic metrics compared to shallower networks. These findings emphasize the limitations of traditional metrics in reflecting real-world clinical efficacy and the necessity of incorporating task-based metrics into model evaluation.

Our recent study [188] focused on evaluating metrics for segmentation and other tasks such as classification, regression, clustering, and image-to-image translation across Python libraries, R packages, and MATLAB functions. The findings revealed that Accuracy, Precision, and Recall for 2D segmentation, as well as Accuracy for 3D segmentation, are consistent across programming languages, underscoring their reliability for benchmarking segmentation tasks. Recognizing that metrics from various studies may not be directly comparable due to differences in study design, participant characteristics, and evaluation methods is essential.

Computational Complexity:When selecting an appropriate 2D segmentation technique for clinical use, it is important to consider computational complexity and user-friendliness. A comparison of eight DL models (UNet, SegNet, GCN, FCN, DeepLabV3 + , PspNet, TransUNet, and SwinUNet) for lung nodule segmentation in CT scans found that TransUNet, with its transformer-based encoder, had the highest accuracy (DSC: 0.813) but was the slowest. The performance of all models improved with the extraction of the ROI over using entire CT scans, while adding lung masks had little to no benefit and sometimes hindered performance. Thus, for precise segmentation, transformer-based models like TransUNet with ROI extraction are recommended, but traditional CNNs such as UNet or SegNet are better for quicker tasks [32].

The review indicates the prevalence of models used for 2D lung cancer segmentation notably UNet. UNet is particularly effective in medical segmentation, such as lung nodule detection, due to its strong performance with limited data. Although it requires substantial annotated data for optimal results, UNet can deliver accurate segmentation with fewer labeled samples, making it ideal for medical applications. Derivatives like SegNet and Attention UNet further improve boundary refinement and feature capture, though UNet's complexity may limit its clinical use in some settings. [32, 190, 191]. CNNs excel in medical image analysis, providing accurate segmentation and versatility across modalities. However, challenges in generalizability and image-specific adjustments impact their clinical effectiveness [192]. The use of advanced CNN models for lung cancer detection and segmentation has significantly improved precision, efficiency, and clinical relevance, aiding diagnosis and treatment planning. However, CNNs may struggle with capturing long-range dependencies in complex structures, where transformers—especially in dual-path architectures—are emerging as a strong alternative for enhanced 2D medical image segmentation [145]. However, their complexity, data demands, potential for overfitting in small datasets [80], and high computational costs pose challenges [146]. Future research seeks to enhance transformers’ interpretability, reduce dependence on large datasets, and improve efficiency. Integrating transformers with CNNs and utilizing self-supervised learning could further advance medical image segmentation.DL models, especially UNet and its variations, are effective in lung nodule 2D segmentation, benefiting from their ability to learn from limited data and perform well across various datasets. However, their success is heavily dependent on the quality and amount of training data and requires significant computational power, posing challenges in resource-limited settings. Models like ResNet and DenseNet, with their DL, improve detection accuracy but also demand substantial computational resources, limiting their use in less equipped environments. Hybrid approaches like Mask R-CNN with K-means and Fuzzy C-Means segmentation offer alternatives that require fewer labeled images or provide computational efficiency, respectively, yet they compromise accuracy or computational demands compared to full DL methods. Despite the high accuracy and adaptability of DL models in lung cancer segmentation, their need for extensive computational resources, large, and annotated datasets, and concern over interpretability and transparency in clinical decisions highlight crucial areas for advancement to make these models more clinically viable and understandable.

This study identifies biases in 2D segmentation models for lung cancer, with a focus on UNet variants. UNet has rapidly emerged as a crucial model for medical image segmentation, thanks to its exceptional performance and distinctive architecture.UNet models are extensively utilized in 2D medical image segmentation tasks, such as identifying lung cancer in CT scans. Their encoder-decoder architecture, complemented by skip connections, adeptly captures both low-level and high-level image features, essential for precise segmentation of intricate regions like lung nodules. Despite their strengths, including straightforward implementation and suitability for processing 2D CT slices with lower computational demands compared to 3D models, UNet faces several challenges that require attention for enhanced clinical effectiveness. Key limitations include class imbalance in medical datasets, which affects model accuracy, especially in differentiating benign from malignant nodules. Although data augmentation is a common remedy, its adoption is not universal, with only 63% of studies employing this technique. This suggests a need for more sophisticated data balancing methods. Additionally, the models often fail to accurately segment smaller lesions due to the limited receptive field of the convolutional layers.Another concern is the stability of UNet models; only a small fraction of studies assess model performance consistency over different initializations, highlighting a gap in robustness evaluation. This is crucial for ensuring that models do not overfit and perform well across various training conditions. Furthermore, while UNet's adaptability is notable, its generalizability across diverse datasets remains questionable. Many studies lack external validation, particularly on non-public datasets, which limits their applicability in real-world settings.To address these issues, future research should focus on more effective strategies for managing class imbalance and enhancing small lesion detection. Enhancing model stability through comprehensive evaluations and extending external validation efforts will improve reliability and adaptability in clinical environmentsIn the review of 2D segmentation for lung cancer CT images, several methodological insights emerge. Only 22% of UNet model studies use non-public datasets for external validation, potentially limiting the generalizability of the findings. Furthermore, while 89% of the studies include rigorous analyses such as ablation studies or comparisons with state-of-the-art models, 25% highlight top metrics without statistical validation. This lack of statistical rigor could lead to misleading conclusions, underscoring the need for more thorough validation in segmentation research.

Research Gaps and Limitations. Usability, Reusability, Reproducibility, and Accessibility (URRA) have become increasingly vital in the scientific community, especially within medical imaging research [66, 193–201]. Our review of studies from 2016 to 2024 highlights ongoing URRA-related challenges, including computational complexity, training time, and robustness to data variability. For example, a 2016 Nature survey [202] found that over 70% of researchers failed to reproduce others' experiments and more than half couldn't replicate their own. Subsequent studies [203–208] consistently reported low replication rates, with only 64–85% of results successfully replicated and many studies struggling with issues related to computational complexity and training time. Additional surveys [209] echoed these concerns, emphasizing how robustness to data variability and other factors contribute to a growing reproducibility crisis that hampers the timely translation of techniques into clinical applications [210]. However, this review did not encompass practical clinical considerations such as computational complexity, training time, and robustness to data variability. Furthermore, future studies should focus on evaluating reproducibility and reusability by re-implementing and comparing previous research using the same or various private and public datasets.

The paper provides a comprehensive review of the progress in computational methods for lung cancer segmentation, particularly emphasizing the superiority of DL models in terms of accuracy and flexibility. Despite these advances, the review identifies several unresolved challenges that need attention for further research. A critical issue highlighted is the trade-off between the speed and accuracy of these models, where models designed for quick processing compromise on detailed accuracy, which is crucial for clinical diagnosis. Conversely, highly accurate models require significant computational resources and time, potentially limiting their practical use in clinical settings. The "black box" nature of DL models, due to their complexity and lack of transparency, poses another significant challenge. This opacity affects their acceptance among clinicians, who prefer a transparent decision-making process. Moreover, these models' performance varies across different datasets, raising concerns about their generalizability and effectiveness in diverse clinical environments. In computational lung cancer studies, developing and validating robust segmentation models using CT images faces several key challenges. Datasets often lack diversity in patient demographics, imaging protocols, and disease prevalence, primarily representing North American populations, which hampers model generalizability across varied clinical scenarios. Additionally, many datasets suffer from imbalanced class distributions, such as fewer positive cases and varying nodule sizes, increasing the risk of overfitting and biased performance. Inconsistent and inaccurate annotations, along with incomplete labeling and the absence of contextual clinical information, introduce noise and uncertainty, undermining model accuracy and effectiveness. The scarcity of sufficient external testing datasets with diverse and comprehensive data further restricts thorough validation in real-world settings. Lastly, relying solely on CT imaging neglects the potential benefits of integrating multi-modality data, which could enhance segmentation precision and overall model performance.

Data Variability: A major challenge in deep learning-based segmentation is the variability in datasets, which arises from differences in imaging protocols, patient demographics, tumor types, and imaging quality across different centers. This variability can lead to inconsistent model performance, limiting the model's applicability to diverse patient populations. Furthermore, class imbalance remains a critical issue in medical imaging, especially in the context of lung cancer, where benign lesions are far more common than malignant ones. Inconsistent preprocessing methods (e.g., variations in image normalization, cropping, and resizing) further exacerbate these issues. To address these concerns, it is essential to promote the use of multi-center, diverse datasets that capture a wide range of patient populations, imaging conditions, and disease severities. Data augmentation techniques, such as rotation, scaling, and elastic deformation, should be employed to increase dataset diversity and improve model robustness. Additionally, strategies to address class imbalance, such as oversampling minority classes or using focal loss, could help to balance the impact of benign and malignant lesions during training.To address these gaps, we propose several strategies to enhance dataset quality and diversity. Collecting data from diverse populations and global sources can improve model generalizability. Moreover, employing data augmentation and multi-center collaborations can increase dataset size and diversity, enhancing robustness and reducing overfitting. Furthermore, developing semi-supervised or weakly supervised learning methods can leverage both labeled and unlabeled data, boosting performance with limited annotations. In addition, standardizing annotation quality through consistent protocols will ensure accurate training data. Expanding external testing datasets with diverse samples will facilitate thorough validation, while integrating clinical and radiological data into multi-modality datasets (e.g., PET, MRI) can provide a comprehensive view of lung cancer, enhancing segmentation precision and model performance.

In Supplemental File 1, we provide a detailed table evaluating the most frequently used models, including UNet and its variants, alongside other prominent approaches for 2D medical image segmentation. The table outlines the key advantages and disadvantages of each model. Among the advantages, these models generally offer a high DSC, demonstrating strong segmentation accuracy, particularly in complex medical imaging tasks. Their ease of implementation makes them accessible for a wide range of clinical applications, and their adaptability to various 2D segmentation tasks further enhances their versatility, enabling effective segmentation of different types of lesions, including lung cancer. However, these models are not without their challenges. One of the primary disadvantages is their computational complexity, especially in advanced variants, which can make them less suitable for resource-constrained environments. Additionally, these models are highly sensitive to data preprocessing steps, such as image normalization, cropping, and resizing, which can affect their performance. Another limitation is the difficulty in accurately segmenting small nodules, as traditional models often struggle with limited receptive fields and may fail to detect or precisely delineate smaller lesions. Despite these challenges, the table provides valuable insights into the strengths and weaknesses of the most commonly used segmentation models, helping guide future research and clinical adoption.

Innovation and Future Direction. The emergence of novel architectures, such as the DPBET [145] and SW-UNet [80], underscores the evolving landscape of DL models for medical image analysis. These models leverage dual-path architectures and hybrid CNN-Transformer mechanisms to capture both local details and global context, offering improved segmentation accuracy. For instance, Wang et al. [145] introduced the Dual-Path Boundary Enhancement Transformer (DPBET) model, a novel approach for lung nodule segmentation that combines global (CNN-Transformer) and edge (DASample) pathways. This dual-path network, incorporating a Cascade-Axial-Prune Transformer (CAP-Trans) for global feature extraction and a boundary enhancement mechanism for edge detection, improves both the recognition of broad contextual information and the precise delineation of nodule boundaries. This addresses the common challenge in traditional CNN-based models, which struggle to balance capturing local details and global context, particularly due to their limited receptive fields. The integration of the DASample module further enhances boundary awareness, which is critical for accurate nodule segmentation in medical imaging. These recent advancements in lung nodule segmentation demonstrate the power of hybrid CNN-Transformer models and dual-path architecture. By effectively combining local detail extraction with global context modeling, these models address the limitations of traditional approaches and provide more accurate and efficient solutions for medical image segmentation. The application of transformers allows for better handling of large-scale contextual information, improving the precision of lung nodule detection and segmentation across various models and techniques.

Future work in lung cancer segmentation should focus on integrating multimodal data, such as CT, PET, MRI, and histopathology, to leverage complementary information for enhanced tumor characterization. Extending 2D methods to 3D segmentation can capture spatial context and volumetric features more effectively, while semi-supervised and unsupervised learning techniques, including generative adversarial networks and contrastive learning, can address the scarcity of annotated data. Federated learning frameworks can enable collaborative training across institutions while preserving data privacy, and attention mechanisms, such as transformers, can improve focus on critical regions within complex lung structures. Incorporating clinical data like patient history and genetic profiles can enhance personalized models, while explainable AI techniques, such as saliency maps, can make models more interpretable for clinical decision-making. Lightweight, real-time models should be developed for resource-constrained settings and intraoperative use, and longitudinal studies focusing on tumor evolution can aid early detection and monitoring. Domain adaptation techniques are essential to improve generalization across datasets, and synthetic data generation using GANs can address class imbalances and enrich training datasets. Additionally, integrating segmentation outputs with radiomics and post-segmentation analytics can support downstream tasks like tumor classification, risk prediction, and survival analysis, advancing both accuracy and clinical utility in lung cancer management.

Nonetheless, the increased model complexity and the need for extensive dataset validation pose challenges to their widespread adoption. Future research should focus on refining these models, optimizing computational resources, and enhancing their generalizability across diverse medical imaging tasks. To enhance the comparison of various 2D segmentation methods, it is crucial to standardize evaluation methodologies. Understanding how image quality and nodule characteristics influence segmentation performance is essential for improving accuracy. Additionally, exploring the fusion of 2D techniques with other approaches such as 3D segmentation or ML can lead to increased robustness and precision. The integration of 2D and 3D segmentation methods is highly beneficial in medical imaging, enhancing accuracy, precision, and contextual awareness. While 2D segmentation provides detailed, slice-level analysis, it can miss broader spatial relationships, which 3D segmentation captures across entire volumes. Combining both methods reduces segmentation errors, improves efficiency, and handles complex anatomical structures more effectively.. The paper is optimistic about the potential of advanced DL models like Long Short-Term Memory (LSTM) networks and GANs in addressing current limitations and advancing segmentation performance. The continuous development of DL models, along with considerations for ethical and computational issues, is crucial for fully realizing the potential of these technologies in clinical settings [148].

In the context of 2D segmentation analysis, the pivotal insight gleaned from the literature review underscores the imperative to prioritize semi-supervised and unsupervised models. This strategic emphasis stems from the inherent complexities associated with image labeling, the limited availability of data, and the critical role played by radiologists in the segmentation process. The research findings advocate for a shift towards leveraging semi-supervised and unsupervised approaches as viable solutions to mitigate the challenges posed by labor-intensive image annotation, data scarcity, and the need for expert input in segmentation tasks. By embracing these advanced modeling techniques, researchers and practitioners can enhance the efficiency and effectiveness of 2D segmentation processes, ultimately advancing the field towards more robust and accurate segmentation outcomes.

Reusability and reproducibility are critical in scientific research, especially in medical imaging. Recent studies reveal challenges such as high computational demands, long training times, and the need for robustness across diverse datasets. Many researchers struggle to replicate experiments, leading to low replication success rates. Data variability and limited resources contribute to a reproducibility crisis, delaying the clinical adoption of new techniques. Future research should prioritize evaluating reproducibility and reusability by re-implementing and comparing previous studies using consistent or various private and public datasets.

In clinical settings, sensitivity and specificity directly influence diagnostic accuracy and patient outcomes, making their optimization critical. Meanwhile, relying solely on task-agnostic metrics may lead to overestimation of model performance and hinder clinical translation. Therefore, we advocate for a balanced approach, integrating standardized task-agnostic and task-based evaluations, as suggested by Liu et al. [189], to ensure robust and clinically relevant model assessments. These findings reinforce the importance of designing evaluation frameworks that align with real-world clinical needs, promoting the development of reliable and impactful diagnostic tools in lung cancer detection. This dual approach not only addresses current gaps in model evaluation but also ensures better translation of research into clinical practice, ultimately benefiting patient care. Moreover, broadening the scope of evaluation metrics to include geometric and uncertainty-based measures, such as the Hausdorff Distance and Free-Response Receiver Operating Characteristics (FROC), will provide a more nuanced understanding of model performance and clinical utility.

Conclusions

This review has aimed to capture the current landscape of 2D segmentation techniques for lung cancer in CT images, underscoring the pivotal role of DL models in revolutionizing this domain. Our comprehensive analysis reveals a field at the nexus of rapid technological advancements and enduring challenges, where the efficiency and adaptability of 2D models shines, albeit not without limitations. The emergence of hybrid models and innovative architecture heralds a promising pathway to surmount these barriers, highlighting the imperative for sustained research and development. To harness the full spectrum of DL's capabilities in medical imaging, future endeavors must prioritize technological innovation and computational efficiency. Continued pursuit of these factors will be crucial in advancing the field towards more accurate, accessible, and reliable diagnostic tools for lung cancer, ultimately enhancing patient care and outcomes in oncology. Current research on 2D segmentation UNet models for lung cancer highlights progress in transparency and preprocessing but reveals gaps in stability, cross-validation, and evaluation metrics. Addressing these gaps and socioeconomic factors is crucial for developing robust, equitable models applicable across diverse clinical settings.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

Conceptualization, M.S and S.M; data analysis, Zh.S, S.M; data gathering and organizing, Zh.S,S.M, F.Ph, A.F and A.M; writing—original draft preparation, S.M, and M.S; writing—review and editing, M.S, A.R, I.H, R.Y; visualization, M.S, A.R, I.H, R.Y; supervision M.S; All authors have read and agreed to the published version of the manuscript.

Funding

This study was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant RGPIN-2019–06467, UBC Department of Radiology 2023–25 AI Fund.

Data Availability

As this is a review paper, no primary data were collected or analyzed. All relevant studies have been cited appropriately in the main text.

Declarations

This article undertakes a systematic review guided by the Preferred Reporting Items for Systematic Reviews and Me-ta-Analysis (PRISMA) framework.

Competing Interests

The authors have no relevant financial or non-financial interests to disclose.