Narrative review: the research advances of artificial intelligence in the prediction of pulmonary nodule growth

Abstract

Background and Objective

Lung cancer is a leading cause of cancer-related mortality worldwide, and dynamic computed tomography (CT) monitoring of pulmonary nodules is central to early detection. However, frequent follow-up scans increase radiation exposure and strain medical resources. This narrative review aims to summarize current artificial intelligence (AI)-based approaches for predicting pulmonary nodule growth on CT, compare model performance, and discuss key challenges and future directions for clinical translation.

Methods

We synthesized peer-reviewed evidence on CT-based volumetric, radiomics, radiogenomics, and machine learning (ML) models related to pulmonary nodule growth or growth-related endpoints. Articles were identified through database searches, reference list screening, and guideline consultation. Eligible publications included human original studies, reviews, and meta-analyses. Conference abstracts, editorials, letters, animal or phantom studies were excluded. No language restrictions were applied.

Key Content and Findings

Existing AI-based growth prediction studies generally follow three workflows: baseline CT models estimating future growth risk, longitudinal CT models modeling voxel- or pixel-wise growth, and models based growth-related surrogates [such as volume or mass doubling time (MDT), invasiveness, or stage shift]. Across these workflows, AI-driven volumetric and radiomic features detect nodule growth earlier and more consistently than simple diameter measurements and better discriminate fast-growing from indolent nodules. At the same time, emerging explainable AI frameworks help identify influential features, potentially improving trust and adoption. However, practical challenges remain, including picture archiving and communication systems (PACS) integration and balancing sensitivity with specificity, overdiagnosis, and false progression. Most models are retrospective and single-center, use heterogeneous protocols and non-standardized growth definitions, and lack external or prospective validation, limiting generalizability.

Conclusions

AI, particularly deep learning (DL) combined with quantitative radiomics and radiogenomics, shows promise for noninvasive prediction of pulmonary nodule growth. Future work should focus on multicenter prospective validation, standardized growth endpoints, low-dose protocols, multimodal data integration, and explainable, federated, generative AI to improve robustness, transparency, and data privacy. In addition, seamless PACS integration and explicit balancing of sensitivity, specificity and overdiagnosis are essential. Ultimately, validated AI models may enable more accurate, personalized surveillance while reducing radiation exposure and resource burden.

Introduction

Background

Lung cancer is one of the leading causes of cancer-related deaths worldwide (1). Early diagnosis can significantly improve patient prognosis. In computed tomography (CT) screening, pulmonary nodules (diameter <3 cm) are the main manifestation of early-stage lung cancer, but approximately 95% of these nodules are benign and do not require intervention (2). Current guidelines recommend monitoring nodule growth through follow-up to assess the risk of malignancy, but multiple CT scans increase radiation exposure and medical burden (3,4). Therefore, predicting the growth trend of nodules with the baseline CT has become a core issue in optimizing lung cancer screening strategies.

The risk of lung cancer increases exponentially with the size of the nodules (5,6), and the diameter and growth rate of the nodules are important predictive factors (7). Traditional assessment is based on changes in two-dimensional diameter, for example, the Lung CT Screening Reporting and Data System (Lung-RADS) guideline defines growth as an increase in diameter of more than 1.5 mm (8). However, the diameter-based standard is not reliable for smaller nodules (9), and the volume doubling time (VDT) has gradually become a core indicator (10). The lung nodule management strategy based on VDT, formulated according to the Nederlands-Leuvens Longkanker Screenings Onderzoek (NELSON) study (11), suggests follow-up at corresponding intervals when VDT is 400 to 600 days or more than 600 days, and further intervention when VDT is less than 400 days. This nodule management protocol was validated through Jiang’s systematic review and meta-analysis (12). For subsolid nodules (SSNs), they proposed a more lenient VDT threshold—600 days for partially solid nodules and 800 days for pure ground-glass nodules (pGGNs)—thereby optimizing screening strategies and improving early lung cancer management. In addition, the mass doubling time (MDT), which combines volume and density, may be superior to VDT due to its higher sensitivity and repeatability (13). One study has also found that density increase is more common in early-growing nodules than is volume increase, and growing nodules grow faster (14). These findings highlight the limitations of traditional assessment methods and the urgent need for more precise quantitative analysis techniques.

Radiomics, by extracting high-throughput features from medical images and combining them with machine learning (ML) analysis, has been widely applied in tumor diagnosis and prognosis assessment (15-17). In recent years, commercial artificial intelligence (AI)-assisted diagnostic software for pulmonary nodules has been extensively used in clinical practice. However, it is still limited to the detection of nodules and the initial prediction of their benign or malignant nature. Predictions regarding their growth trends or patterns are still rarely reported (18). Although deep learning (DL) has further promoted the development of automatic lung nodule segmentation and growth prediction models, existing studies still face challenges in data heterogeneity, model generalization, and clinical translation (19). The precise calculation of VDT and MDT relies on high-quality nodule segmentation and dynamic follow-up data, while the performance of current AI models in complex nodule morphology and heterogeneous data still needs improvement. We present this article to systematically review the technical advances in lung nodule growth prediction, to propose future research directions, and to provide theoretical support for optimizing lung cancer screening strategies and facilitating the clinical translation of AI technologies. We present this article in accordance with the Narrative Review reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1580/rc).

Methods

Information used to write this paper was collected from the sources listed in Table 1 and Table S1.

Table 1. The search strategy summary.

Items	Specification
Date of search	20 May 2025, 2 December 2025, 3 December 2025, and 4 December 2025
Databases and other sources searched	PubMed, Google Scholar, Embase, IEEE Xplore
Search terms used	(Pulmonary nodule OR lung nodule OR ground glass nodule OR part-solid nodule) AND (growth OR grow OR volume OR mass OR volume doubling time OR mass doubling time OR progression OR temporal OR longitudinal change OR interval change OR evolution OR growth trend OR growth pattern) AND (AI OR artificial intelligence OR radiomics OR convolutional neural network OR deep learning)
Timeframe	1/2020–12/2025
Inclusion and exclusion criteria	We included peer-reviewed original research, reviews, and meta-analyses in human subjects that evaluated CT-based volumetric, radiomics, radiogenomics, or AI/machine-learning approaches relevant to pulmonary nodule growth or its quantitative surrogates (e.g., VDT, MDT, invasiveness, stage shift). No language restrictions were applied. Publications from database inception to November 2025 were considered; AI-based growth prediction studies meeting the inclusion criteria were published between 2020 and 2025
Selection process	W.J. and L.F. conducted the selection process independently and separately, and a consensus was obtained through discussions among all group members

AI, artificial intelligence; CT, computed tomography; MDT, mass doubling time; VDT, volume doubling time.

Discussion

Radiomics and DL

Traditional radiomics

The traditional radiomics consists of the following steps: (I) image acquisition: obtain high-quality CT images with a slice thickness of ≤1 mm to minimize potential volume measurement errors. (II) Nodule segmentation: manually, semi-automatically, or automatically delineate the region of interest (ROI) for the nodule, focusing on either the maximum cross-sectional area or the three-dimensional volume. (III) Feature extraction and selection: extract morphological features (e.g., diameter, volume), texture features (e.g., gray-level co-occurrence matrix), and high-order features (e.g., wavelet transform), followed by the selection of significant features using methods such as least absolute shrinkage and selection operator (LASSO) regression or random forest (RF). (IV) Model construction: integrate clinical parameters (e.g., age, smoking history) with imaging features to construct predictive models, such as RF or support vector machine (SVM), and evaluate their performance using receiver operating characteristic (ROC) curves. (V) Clinical application: the model predicts the growth trend, pattern, or probability of nodules, thereby assisting in the formulation of individualized follow-up strategies.

DL

DL primarily enables the automatic extraction of image features through convolutional neural networks (CNNs), thereby overcoming the dependency of traditional methods on manual feature engineering. Key applications include: (I) nodule segmentation: The U-Net architecture facilitates high-precision three-dimensional segmentation of pulmonary nodules, reducing inconsistencies in manual annotations. (II) Growth prediction: Recurrent neural networks are trained using sequential CT images to capture temporal changes in nodule volume and density. (III) Multi-task learning: Simultaneous prediction of nodule malignancy, genetic mutations such as epidermal growth factor receptor (EGFR), and growth rate enhances the overall performance of the model.

Typical workflow for pulmonary nodule growth

Alignment strategies for baseline and follow-up scans in longitudinal assessment and the application of deformable registration algorithms (DRAs)

Alignment of baseline and follow-up scans

In the longitudinal assessment of pulmonary nodules, accurate alignment between baseline and follow-up scans is a prerequisite for dynamic quantitative analysis and malignant risk evaluation, yet its core obstacles arise from multi-dimensional factors (20-23).

Firstly, physiological movements such as respiratory phase differences (causing up to 38.75% lung volume change and 20.21% maximum volume variation in SSNs), heartbeat and postural shifts induce nodule misalignment, with pGGNs show more significant deformation than part-solid nodules (PSNs) due to incomplete alveolar collapse; additionally, intrinsic nodule characteristics like small size (<8.4 mm), volume reduction, size change rate and heterogeneous deformation between pGGN and PSN increase tracking failure risk; furthermore, disparities in CT scanners and reconstruction parameters across centers lead to inconsistent image quality, which conventional rigid registration cannot address; finally, manual alignment by radiologists is labor-intensive, time-consuming and variable, failing to meet large-scale screening and long-term follow-up needs.

Necessity of the application of DRA

By establishing pixel-level nonlinear spatial mapping, DRA address the issues of large deformation and local distortion of lung tissue that cannot be handled by traditional rigid registration, thus becoming a core supporting technology for longitudinal assessment.

First and foremost, DRA can improve alignment accuracy. By capturing the spatial displacement caused by respiratory movement and organ deformation, they achieve precise matching of anatomical structures between baseline and follow-up scans, reducing the measurement error of nodule size [e.g., the target registration error (TRE) is reduced to 1.44±1.24 mm] and enhancing the consistency of morphological features (22).

Secondly, DRA can empower automated analysis: registration serves as the foundation for automatic nodule tracking, dynamic quantification (i.e., changes in volume/density), and longitudinal modeling of malignant risk. For instance, the registration algorithm based on U-net achieves an automatic nodule tracking success rate of 94% (21).

Thirdly, DRA can reduce clinical burden: automated registration replaces manual operations, shortening the analysis time (e.g., the registration time of the multi-scale residual algorithm is less than 2 seconds). Meanwhile, it reduces inter-observer variability and improves the consistency of assessment (22).

Finally, in multi-center studies, by standardizing image spatial mapping, DRA can alleviate the systematic biases caused by different equipment and scanning parameters, and improve the generalization of research results (23).

Key technologies of current DRA

Current research focuses on addressing issues such as large deformation, uncertainty, and multi-scale features. Representative algorithms and their effects are as follows, with core technical parameters and performance verified by clinical datasets:

1. DL-based registration frameworks
Multi-scale residual network (22): it captures multi-scale deformation features of lung tissue through a multi-resolution self-attention fusion module, and optimizes feature extraction by combining feature-corrected skip connections. On the Dir-lab dataset, the TRE reaches 1.44±1.24 mm, the voxel folding rate is less than 0.1%, and the registration time is less than 2 seconds.
Stochastic decomposition strategy (24): it randomly decomposes large deformation fields into small deformation components, and enhances the registration stability of large deformation regions through intermediate proxy image supervised training. The TRE is reduced by 12.3% compared with traditional methods, and the Dice coefficient reaches 0.9837 on the Dir-lab dataset.
U-net-like models (21): nodule tracking is achieved by predicting deformation vector fields. Among 49 nodules from 40 patients and 368 follow-up CT scans, the success rate of single assessment is 94%, and the success rate of nodule-level assessment is 78%.

2. Semi-automation and cross-modal extension
Cross-modal registration (25): deformable registration between magnetic resonance imaging (MRI) and low-dose CT (LDCT) enables radiation-free follow-up. The consistency of Lung-RADS classification is κ=0.86–0.88, and the consistency of nodule growth assessment is κ=0.88–1.0, providing a new solution for long-term monitoring of high-risk groups.

Classification and comparison of three typical analysis workflows

Workflow 1—traditional radiomics analysis pathway

Workflow 1 can be outlined as predicting lung nodule growth by extracting valuable features from the image, followed by feature dimensionality reduction, modeling with a classifier based on the selected features. This radiomics process is the model most commonly used in the field of disease analysis. This process usually involves manual extraction operations of lesion features. For example, Zhao et al. (26) successfully predicted the invasiveness of early-stage lung adenocarcinoma across multiple centers using semantic features from radiomics and CT scans, achieving preoperative non-invasive prediction. This provides surgeons with valuable references for selecting appropriate treatment strategies and surgical plans. Zhang et al. (13) further investigated the instability of ground-glass nodules (GGNs)-type lung adenocarcinoma by evaluating the dynamic changes in lesions through MDT. They found that the model achieved optimal predictive performance when the MDT duration reached 813 days. These findings demonstrate that radiomics features can effectively identify unstable nodules with rapid growth potential, offering a basis for personalized follow-up strategies. For clinical applications, the introduction of the radiomics model breaks through the traditional methods of feature selection such as radiological features and clinical features, and brings a new perspective to the accurate prediction of nodal growth.

Workflow 2—DL-based end-to-end prediction

Workflow 2 is summarized as follows: DL algorithms automatically perform image processing based on image information, such as image segmentation, classification and other tasks, and ultimately complete the model construction, without human intervention in the period. Li et al. (27) proposed a novel pixel-based nodule growth prediction network that achieves precise growth prediction by forecasting pixel-level variations in nodule masks. The model outperforms existing methods such as 3D Residual U-Net (ResUNet) and parameterized Gompertz-guided Morphological AutoEncoder (GM-AE) in metrics like dice similarity coefficient (Dice) and intersection over union (IoU), particularly excelling in the growth-only test set (a subset of nodules with volume increase exceeding 50%). This demonstrates clinical value for malignancy assessment and treatment planning. Agrawal et al. (28) developed a DL model-based nodule detection and segmentation that enables automated tracking of nodule progression on low-dose CT scans through precise boundary segmentation and registration of follow-up images. The most important feature of Workflow 2 compared to Workflow 1 is that it increases efficiency and reduces observer discrepancies that may result from extensive human labeling.

Workflow 3—combined model analysis pathway

Workflow 3 focuses on lung nodule growth prediction using a combined model consisting of clinical features, radiomics features, and other features. Unlike the previous three workflows, this process fuses multimodal data for evaluation, and its model performance is substantially improved. For example, Luo et al. (29) constructed a combined model combining radiomics features and five clinical predictors, which outperformed a conventional radiomics model in identifying high-risk growth patterns in invasive lung adenocarcinoma [area under the curve (AUC): 0.923 vs. 0.888]. Another cohort study (30) found that a combined model constructed on the basis of relevant radiomics features and clinical biomarkers at preoperative time points was effective in predicting SSN invasiveness. This is significant for the early intervention of aggressive nodules and also helps to reduce the negative impact of unnecessary clinical interventions (Figure 1).

Figure 1.

Schematic diagram of the three workflows.

Current status of predicting lung nodule growth

Prediction of nodule growth risk

As for the risk factors of nodule growth, several studies have revealed key predictive indicators. In the univariate analysis of two-dimensional cross sections (31), nodule area, perimeter, diameter, linear mass density (LMD), circularity and solidity were significantly associated with growth, whereas skewness of CT attenuation and LMD were identified as independent predictors of nonsolid nodules growth in multivariate analysis. Clinical visualization parameters and three-dimensional reconstruction technology expanded the evaluation dimension (32). Yuan et al. (33) demonstrated that, through multivariate logistic regression analysis, nodule type (PSN), irregular morphology, and pleural traction were identified as independent predictors of growth in SSN. A meta-analysis by Wu et al. (34) (including 2,898 SSN). indicated that the initial nodule volume was the strongest predictor of growth incidence and time to growth. Furthermore, a retrospective study by Tang et al. (35) using VDT to predict interval growth of SSN revealed that a VDT ≤400 days was a closely associated predictor of aggressive growth behavior with a stage shift.

Based on the clarification of growth factors, the AI technique further reveals the growth risk of nodules. The prediction model developed using the Wasserstein generative adversarial network framework analyzes growth patterns in LDCT images, and accurately predicted lung cancer risk within 1 year for nodules, confirming the feasibility of using Workflow 2 to improve clinical management of lung nodule screening, thereby facilitating the early diagnosis of lung cancer (36). Similarly, Wang et al. (37) enhanced the growth prediction efficacy of pulmonary nodules with the assistance of Workflow 2. Their developed lung nodule growth network (LNGNet) model incorporates an adaptive temporal scaling (ATS) module to dynamically adjust temporal scales, thereby enhancing the model’s robustness against temporal heterogeneity. By dynamically updating the baseline nodule texture weights, it mitigates the interference of varying CT image backgrounds across different time points on the generated results. The final performance shows a 5% improvement in volume prediction AUC and a 7.2% improvement in mass prediction AUC compared to the baseline. Additionally, the Dice coefficient for shape generation increased by 2.89%, providing a high-precision and practical tool for nodule growth assessment.

In terms of variable integration, Li et al. (38) systematically compared three models predicting GGN invasiveness based on follow-up data changes, finding that the radiomics nomogram constructed by a combined model integrating morphological features, quantitative parameters, and follow-up growth rates demonstrated the highest diagnostic value (AUC =0.932). Workflow 3 further improves the prediction of growth trends in solid pulmonary nodules, demonstrating significant clinical value (39). For GGN, the radiomics model developed by Jin et al. (40) accurately predicts the progression status (absorption or persistence) of nodules on initial CT scans, outperforming experienced radiologists. When combined with clinical factors, the model achieves optimal predictive performance (AUC =0.959), facilitating optimized clinical follow-up management of GGN and reducing unnecessary examinations. For SSN, Chen et al. (41) found that the integrated radiomics-DL model demonstrated the best performance (AUC =0.926). Yuan et al. (33) further applied Workflow 3 to subpleural SSN, demonstrating that the imaging-radiomics composite model could effectively predict nodule growth and stability. This finding highlights the potential clinical value of future integrated models.

Prediction of growth rate and growth pattern of pulmonary nodules

In traditional quantitative analysis, the combination of VDT and density metrics reveals that malignant nodules often exhibit an accelerated, faster-than-exponential growth pattern (5). The dynamic characterization of growth patterns and advances in quantitative assessment techniques have provided new perspectives for accurate prediction. Qi’s team (42,43) constructed a growth assessment model through Workflow 2 and found that persistent pGGN mainly showed indolent exponential growth by DL-assisted segmentation, whereas SSN with pathology of invasive adenocarcinoma (IAC) grew faster and the risk of growth was higher with a larger initial volume. A study (12) indicates that adenocarcinomas demonstrate a higher proportion of indolent growth (68.9%), whereas squamous cell carcinoma, small cell carcinoma, and other histological types show extremely low rates of indolent growth. The solid components, tumor size, and stage of the tumor are closely associated with growth rate, a finding consistent with most studies. Notably, their research found no correlation between gender, age, and smoking volume with nodule growth rate. With the development of computer science, CT texture analysis based on gray level cooccurrence matrix provides new tools for revealing tumor heterogeneity. This method, which indirectly reflected the heterogeneity of cells at the molecular level through the pixel distribution characteristics (44), not only deepened the understanding of the biological behavior of nodules, but also provided dual potential for noninvasive assessment of malignant potential and dynamic monitoring.

Prediction of VDT/MDT

VDT is a key indicator for evaluating lung nodule growth. A systematic review and meta-analysis (12) incorporating 33 studies (3,959 patients) revealed that nodules with a VDT >400 days are mostly indolent, accounting for 34.9% of all lung cancer cases and 68.9% of adenocarcinoma cases in the included studies. Currently, multiple studies have developed innovative growth rate prediction models based on VDT and MDT, and the relevant achievements provide important technical support for clinical practice.

Jiang et al. (45) developed an AI-based model for early malignant nodule detection and a VDT assessment system. The study analyzed 710 lung cancer screening patients, demonstrating lower error rates and superior recognition accuracy compared to expert evaluations, thereby providing technical support for personalized treatment. Huang et al. (46) focused on the performance of different ML methods for predicting the VDT of GGN, compared seven ML algorithms and selected the neural network (NNet) model with the highest robustness against data perturbations. The single ML model achieved an accuracy of 0.756 on the external verification set, which further showed the important application prospect of ML in predicting VDT. Han et al. (47) adopted a 3D ResNet model to predict VDT ranges (with 400 days as the threshold). In a cohort of 734 patients with solid nodules, the designed Workflow 2 achieved an accuracy of 81.55% in predicting VDT of pulmonary solid nodules, and the prediction could be completed in only 5–7 seconds, significantly reducing manual image review time and improving work efficiency.

However, the impact of density changes on growth is difficult to assess in the above-mentioned models and cannot fully reflect the heterogeneity of nodule growth. There may be an underestimation risk for GGN mainly characterized by density alterations. In contrast, the MDT, which calculates mass changes by combining nodule volume and CT values, can more comprehensively reflect the dual characteristics of nodule growth in terms of morphology and density, and has been proven to be more sensitive than VDT in identifying the invasiveness and instability of GGN.

It is worth noting that multiple studies have clearly demonstrated the technical advantages and clinical value of MDT over VDT (5,13,48-50). Zhang et al. (13) conducted a study on 298 cases of GGN-type lung adenocarcinoma and found that MDT was closely related to the invasiveness of the nodules. The radiomics model (B1) based on MDT achieved an AUC of 0.89 in predicting nodule instability, significantly outperforming traditional volume assessment. When the MDT threshold was set at 813 days, the model’s specificity (0.83) and accuracy (0.78) were optimal. Sekine et al. (49) confirmed that the median MDT (789 days) was significantly shorter than VDT (1,000 days), and the MDT of PSNs (588 days) was significantly shorter than that of pure GGN (961 days), which could more accurately identify occult growth. He et al. (50) found that MDT was significantly negatively correlated with Ki-67 expression, and Ki-67 is a core indicator reflecting tumor proliferation activity, suggesting that MDT is more closely associated with the biological behavior of nodules. Additionally, MDT can effectively avoid the misjudgment of VDT for nodules that are “volume stable but density increased”, and has shown unique value in the early malignant risk stratification of GGN (13,49).

The above research results indicate that radiomics and DL techniques have been widely applied in the prediction of VDT/MDT for pulmonary nodules. MDT, with its comprehensive assessment of the “volume-density” dual dimensions, has gradually become a superior growth evaluation index, promoting the clinical transition from the single VDT standard to the combined assessment of VDT and MDT. With the continuous update and iteration of related technologies and the optimization and upgrading of algorithms, advanced models integrating MDT are expected to be more accurately applied in real medical scenarios, providing efficient support for individualized diagnosis and treatment of pulmonary nodules. Table 2 summarizes the prediction results of lung nodule growth based on different workflows.

Table 2. Summary of studies predicting nodal growth based on four workflows.

Study	Year	Imaging modality	Lesion	Design	No. of patients	Feature type	Developed model	AUC*
Zhang et al. (13)	2024	CT	Lung adenocarcinoma	Retrospective single center	249	Radiomics features	SVM	0.89
Zhao et al. (26)	2024	CT	Lung adenocarcinoma	Retrospective multi-center	572	Radiomics features and CT semantic characteristics	LR	0.911
Li et al. (27)	2024	CT	Pulmonary nodule	Retrospective single center	349	Raw image data	DCNN	–
Agrawal et al. (28)	2023	CT	Pulmonary nodule	Prospective single center	110	Raw image data	DL	–
Luo et al. (29)	2025	CT and PET	Lung adenocarcinoma	Retrospective single center	203	Radiomics and clinical features	LR	0.923
Singh et al. (30)	2025	CT	SSN	Retrospective single center	186	Delta radiomics, delta volume and clinical features	LR	0.69
Yuan et al. (33)	2025	CT	SSN	Retrospective single center	454	Radiomics features	LR	0.896
Wang et al. (36)	2024	LDCT	Pulmonary nodule	Retrospective single center	1,226	Raw image data	GP-WGAN	0.862
Wang et al. (37)	2024	CT	Pulmonary nodule	Retrospective single center	103	Imaging features	LNGNet	0.554
Li et al. (38)	2024	CT	Lung adenocarcinoma	Retrospective single center	687	Quantitative imaging features and clinical features	LR	0.932
Liang et al. (39)	2024	PET-CT	Pulmonary nodule	Retrospective single center	300	Radiomics and clinical features	LR	0.886
Jin et al. (40)	2026	CT	GGN	Retrospective multi-center	672	Clinical information and radiomic features	linear SVC	0.969
Chen et al. (41)	2025	CT	SSN	Retrospective multi-center	353	Radiomics features and raw image data	RF	0.926
Qi et al. (42)	2020	CT	pGGN	Retrospective single center	110	Quantitative imaging features	CNN	–
Qi et al. (43)	2021	CT	SSN	Retrospective single center	95	Quantitative imaging and radiologic features	CNN	0.819
Jiang et al. (45)	2026	CT	Pulmonary nodule	Retrospective multi-center	710	Quantitative imaging features	AI software	–
Huang et al. (46)	2023	CT	GGN	Retrospective multi-center	172	Radiomics features	NNet	0.896
Han et al. (47)	2024	CT	Solid pulmonary nodule	Retrospective single center	734	Raw image data and clinical information and imaging features	CNN	–
Xiong et al. (48)	2024	CT	GGN	Retrospective single center	755	Clinical information and quantitative imaging and radiologic features	CNN	0.883

*, in this table, the AUC is from the top-performed model of the validation or test set (where available). AI, artificial intelligence; AUC, area under the curve; CNN, convolution neural network; CT, computed tomography; DCNN, deep convolution neural network; DL, deep learning; GGN, ground-glass nodules; GP-WGAN, Wasserstein Generative Adversarial Network framework; LDCT, low-dose computed tomography; LNGNet, lung nodule growth network; LR, logistic regression; NNet, neural network; PET, positron emission tomography; PET-CT, positron emission tomography-computed tomography; pGGN, pure ground-glass nodules; RF, random forest; SSN, subsolid nodules; SVC, support vector classification; SVM, support vector machine.

The association between molecular mechanisms and the growth of pulmonary nodules

A previous study has shown that radiomics features can effectively characterize the heterogeneity of the tumor microenvironment and serve as non-invasive biomarkers to predict the growth of lung adenocarcinoma (51). The field of radiogenomics has also demonstrated great potential by establishing quantitative associations between radiomics features and molecular variations, opening up new avenues for non-invasive prediction of tumor genotypes. Table 3 summarizes the studies on the correlation between pulmonary nodules and molecular characteristics. Zhang et al. (52) based on preoperative CT image data of 424 lung adenocarcinoma patients, significantly improved the accuracy of predicting EGFR mutations by constructing a combined model of multi-phase radiomics and clinical features, with an AUC of 0.927. Mahajan et al. (53) also demonstrated high accuracy in predicting EGFR mutations (AUC =0.88), further confirming the indirect mapping ability of radiomics features to molecular variations. Compared with traditional CT features, the recently proposed radiomics based on topological images derived from CT images (54) has shown higher and more robust predictive ability in predicting EGFR mutations in non-small cell lung cancer (NSCLC).

Table 3. Summary of studies related to pulmonary nodules and molecular characterization.

Study	Year	Imaging Modality	Lesion	Design	No. of patients	Feature type	Developed model	AUC*
Yoon et al. (51)	2020	CT	Lung adenocarcinoma	Retrospective single center	52	Radiomic margin features	Multiple GEEs	–
Zhang et al. (52)	2024	CT	Lung adenocarcinoma	Retrospective multi-center	424	Radiomics and clinical features	XGBoost	0.927
Mahajan et al. (53)	2024	CT	NSCLC	Retrospective multi-center	990	Radiomics and semantic features	SVM	0.90
Kodama et al. (54)	2025	CT	NSCLC	Retrospective multi-center	226	Persistent lifetime features	RF	0.984
Liang et al. (55)	2024	CT	Pulmonary nodule	Retrospective single center	432	Molecular features	Unsupervised hierarchical clustering	–
Li et al. (56)	2025	CT	Lung adenocarcinoma	Retrospective multi-center	1,212	Radiomics features	SVM	–

*, in this table, the AUC is from the top-performed model of the validation or test set (where available). AUC, the area under the curve; CT, computed tomography; GEEs, generalized estimating equations; NSCLC, non-small cell lung cancer; RF, random forest; SVM, support vector machine; XGBoost, eXtreme Gradient Boosting.

It is worth noting that there is a close biological association between genotypes and tumor growth phenotypes. A recent study (55) has revealed the dynamic role of EGFR mutations in the evolution of lung adenocarcinoma: this mutation is not only significantly associated with early tumor growth but also shows an increasing frequency trend during the transition from non-invasive to invasive stages. Additionally, Tumor protein 53 (TP53) mutations are more closely related to invasive lung nodules. This genotype-phenotype correspondence provides theoretical support for the clinical application of radiogenomics: through quantitative analysis of radiomics features, it may be possible in the future to directly predict the growth dynamics of nodules and discover new growth-promoting driver genes, thereby providing a molecular-level scientific basis for formulating individualized intervention strategies. A recent cohort study explored the association between radiomics features of pulmonary nodules and molecular characteristics. Li et al. (56) classified lung adenocarcinoma into four clusters based on radiomics consensus clustering and constructed an SVM classifier to provide biological explanations for the different clinical pathological differences within the clusters through ribonucleic acid (RNA) sequencing. The results showed that cluster 1 represented slow-growing and well-differentiated nodules, while cluster 4 represented advanced cell development. Highly proliferative nodules were classified into clusters 2 and 3.

The importance of explainable AI in predicting pulmonary nodule growth

One of the core challenges in deploying AI models for predicting pulmonary nodule growth is the “black box” nature of these models, which undermines clinical trust and limits interpretability. Explainable AI (XAI) techniques, such as saliency maps, local Interpretable model-agnostic explanations (LIME), gradient-weighted class activation mapping (Grad-CAM) and SHapley Additive exPlanations (SHAP), can reveal which image regions contribute most to model predictions (57). For example, by combining SHAP with feature importance analysis, it is possible to quantitatively assess the relative contributions of radiomic features and clinical variables to the prediction of nodule invasiveness (58). Similarly, visualization-based explanation methods in deep CNNs (such as Grad-CAM and SHAP) have shown that handcrafted features, especially shape descriptors, are strongly associated with malignancy probability, thereby providing clinically meaningful explanations for differentiating benign from malignant nodules (59).

In addition, the explainable artificial neural network model proposed by Ibrahim and colleagues, compared with Shapley-based methods, achieve near-optimal predictive performance while substantially reducing model complexity (60). More recently, GPT-4o has been applied for the first time to longitudinal CT-based dynamic assessment of pulmonary nodules without the need for complex network architectures, which can generate high-quality radiologic diagnostic evidence and thereby improve clinical acceptability relative to conventional DL models (23). Collectively, these methods enhance model transparency and allow clinicians to visually inspect whether the cues emphasized by AI when predicting nodule growth are medically plausible—such as irregular nodule margins, vascular convergence, or focal changes in density.

Notably, a recent systematic analysis of XAI applications in oncology found that CNN account for approximately 31% of the models used, SHAP is the predominant interpretability framework (44.4%), and python is the most frequently used programming language (32.1%). Only 7.4% of studies address security issues (61). These findings provide an important reference for future XAI applications in medical imaging and oncology. Therefore, future research should not only focus on predictive accuracy, but also prioritize interpretability and security, in order to foster clinician trust, facilitate regulatory acceptance, and support the safe integration of AI models into real-world clinical practice.

Challenges, potential solutions and prospectives

Although AI has shown great potential in the prediction of lung nodule growth, current research has the following limitations:

1. Data heterogeneity and insufficient model generalization: most existing studies are based on single-center or small sample datasets. There are significant differences in CT equipment parameters, scanning protocols, and patient populations among different institutions, which leads to a decline in model performance when validated across datasets. Although DL models perform well on specific datasets, they lack large-scale multi-center clinical trial support, limiting their clinical applicability.
Possible solutions: (i) federated learning can be used as a key technology to integrate multi-center data through a distributed training framework, improving model generalization while protecting privacy. For example, Mahajan et al. (53) developed a cross-center DL framework based on CT images that successfully predicted EGFR mutation status, demonstrating the effectiveness of federated learning in heterogeneous data. (ii) Establish a standardized data sharing platform to unify imaging acquisition protocols and feature extraction processes, and introduce dynamic calibration algorithms to reduce the impact of equipment differences on model performance.

2. Beyond data heterogeneity, a major barrier to the clinical practice of AI is integration into real clinical workflows. Most current studies still evaluated AI systems as stand-alone research tools running on exported CT data, with few models embedded into the systems radiologists routinely use—namely picture archiving and communication systems (PACS) or radiology information systems (RIS). This “out-of-workflow” deployment, detached from the diagnostic environment, is likely to slow true clinical translation of AI.
Possible solutions: (i) enable direct, in-workflow visualization of model-derived growth predictions within PACS reading workstations and other online viewers. Results could be presented as vendor-neutral Digital Imaging and Communications in Medicine (DICOM) overlays, growth curves, or risk scores that can be reviewed synchronously with the source images. Existing pulmonary nodule evaluation software already allows seamless retrieval during image interpretation, and growth-prediction tools should be integrated in a similar way. (ii) Ensure that system integration is compatible with established reporting templates, screening programs (e.g., Lung-RADS-based triage pathways), and hospital information technology (IT) infrastructure. This requires close co-development and ongoing collaboration between vendors, radiologists, and IT departments to address cybersecurity, patient safety, maintenance, and monitoring of real-world performance.

3. Another key challenge is overdiagnosis. Many subsolid or indolent nodules exhibit minimal growth over long periods and may never become clinically meaningful. If highly sensitive measurements detect such minor changes and label them as “progression risk”, this may trigger excessive surveillance imaging, invasive procedures and increase patient anxiety.
Possible solutions: (i) future studies should link growth prediction to clinically meaningful outcomes such as stage migration or the need for invasive intervention. (ii) Prospectively evaluate AI-assisted growth prediction within guideline-based management pathways (e.g., Lung-RADS, Fleischner recommendations) and use decision-curve analysis to determine whether AI reduces, rather than amplifies, overdiagnosis. Existing international consensus statements on surveillance intervals for growing nodules provide a useful framework for evaluating the clinical impact of AI-derived “progression” labels.

4. Dynamic monitoring relies on frequent imaging examinations: current growth prediction requires multiple CT follow-ups (e.g., VDT calculation requires at least two scans), increasing the risk of radiation exposure and medical costs for patients.
Possible solutions: (i) generative adversarial networks can synthesize temporal CT images to simulate nodule growth patterns, thereby reducing the number of actual scans (36). (ii) Combine temporal modeling [such as Transformer or long short-term memory (LSTM)] with baseline radiomics features and use time extrapolation techniques to predict growth trends, reducing the reliance on frequent follow-ups.

5. The relationship between radiomics features and molecular mechanisms is unclear: The causal relationship and molecular heterogeneity between radiomics features (such as texture and morphology) and nodule growth dynamics (such as malignant proliferation and gene mutations) limit the accuracy of precise predictions.
Possible solutions: (i) integrate liquid biopsy (such as ctDNA, circulating tumor deoxy RNA) with radiomics to elucidate the association between gene mutations and imaging phenotypes through multimodal data. Liang et al. (55) constructed a genomic-imaging joint map that revealed a significant correlation between EGFR and TP53 mutations and aggressive nodule growth, providing a reference for molecular-oriented prediction models. (ii) Develop radiogenomics combined models, such as combining ctDNA test results with CT texture features, to improve the prediction accuracy of growth dynamics.

Future research still needs to verify the clinical value of the above technologies through multi-center prospective studies and promote the deep integration of AI models with low-radiation monitoring technologies and molecular markers, ultimately optimizing individualized lung cancer screening strategies.

Conclusions

AI has shown significant potential in predicting lung nodule growth through multiple workflows such as radiomics, DL, volume/MDT prediction, and combined models. These tools provide new opportunities for risk stratification and individualized management, but key challenges remain: data heterogeneity limits the generalization ability of the models, dynamic monitoring relies on frequent CT scans, and the correlation between radiomics features and the molecular mechanisms driving nodule growth is still unclear. Future breakthroughs require multi-center prospective validation, combination of low-dose imaging and molecular biomarkers (such as ctDNA), development of federated learning and generative AI technologies, and seamless integration into PACS-based workflows. Equally important is the explicit balancing of increased sensitivity against specificity and overdiagnosis risk, ensuring that AI-based growth prediction improves patient outcomes rather than simply increasing the number of “progressive” nodules flagged. Addressing these issues will be essential for integrating AI into optimized clinical lung cancer screening pathways.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.