Narrative review of the application of artificial intelligence-related technologies in the diagnosis of pulmonary nodules with recommendations for clinical practice and future research

Introduction

Background

Lung cancer is one of the leading causes of cancer-related death worldwide (1), and the prognosis of lung cancer patients strongly depends on the early diagnosis and staging of the tumor. Surgical intervention is currently the main curative treatment for patients with early-stage lung cancer (2). Patients with early-stage resectable disease have a higher survival rate compared to those diagnosed at a metastatic stage. Multiple studies have shown that screening for lung cancer is cost-effective and provides benefits associated with early lung cancer prediction (3,4). The early screening and diagnosis of lung cancer chiefly rely on chest X-rays and computer tomography (CT). Clinicians and radiologists can distinguish benign from malignant pulmonary nodules using CT images, but this practice also increases the workload of medical staff. This contributes to both false positives and false negatives, which, to some extent, reduces the accuracy of early diagnosis and differentiation of pulmonary nodules (5,6). With the advancement of computer technology and statistical science, artificial intelligence (AI) has experienced rapid advancements in recent years. With the progress and deepening of research on computer-aided diagnosis (CAD) and other pulmonary nodule auxiliary diagnosis models, its application in clinical imaging is particularly worthy of attention. Classic AI models, such as convolutional neural networks (CNNs), Residual Neural Network (ResNet), U-Net, etc., can detect various lesions quickly and effectively, and have important clinical application potential in the screening and diagnosis of early-stage lung cancer. For instance, MGI-CNN (Multi-scale Progressive Ensemble CNN) integrates features of different resolutions to reduce false detections (7). Hamidian et al. (8) proposed a CAD system based on three-dimension convolutional neural network (3D-CNN), using three-dimension Fully Convolutional Network (3D-FCN) to mark pulmonary nodules and then conduct recognition. The results verified that using the Lung Image Database Consortium (LIDC-IDRI) dataset of 833 pulmonary nodules, 3D-FCN has increased the detection speed by approximately 800 times compared with the traditional 3D-CNN. Nowadays, AI has become more efficient and accurate in the diagnosis of pulmonary nodules than before. For example, Sha et al. (9) developed a deep learning (DL) model based on ResNet to predict the PD-L1 state. Even when labels of different proportions were randomly closed to simulate the differences between pathologies [the area under the curve (AUC) =0.63–0.77, P≤0.03)], the model remained valid within the PD-L1 cut-off threshold range (AUC =0.67–0.81, P≤0.01). In other respects, Chen et al. (10) used model based on CNN integrated CT image features, seven autoantibodies [such as protein gene product 9.5 (PGP9.5) and G antigen 7 (GAGE7)], and four tumor markers [such as cytokeratin fragment antigen 21-1 (CYFRA21-1)], increasing AUC for the diagnosis of early malignant pulmonary nodules to 0.884, which was superior to the traditional Mayo model.

To sum up, AI has made significant achievements in various fields of pulmonary nodule diagnosis. AI can assist doctors in diagnosis, increasing efficiency and the positive rate. However, is this sufficient? Most models rely on single-center data. Differences in equipment parameters and scanning protocols among different hospitals lead to a decline in model performance (11). When AI misdiagnoses, the legal responsibility is not clearly defined, which hinders clinical application (12). A few models were verified through prospective multicenter trials, and the majority remained at the retrospective research stage (13). The evaluation indicators used in different scientific research, such as AUC and positivity, are difficult to compare the performance of models horizontally (11). This article reviews the clinical application value of AI in the diagnosis of pulmonary nodules and related lung cancer fields, providing clinicians with more comprehensive and detailed diagnosis and treatment suggestions. We present this article in accordance with the Narrative Review reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1512/rc).

Methods

Recent research on the application of AI in pulmonary nodule diagnosis was identified through systematic searches of PubMed (Medline), Web of Science, Cochrane Library, and China National Knowledge Infrastructure (CNKI). The search strategies are summarized in Table 1.

Discussion

Application of AI in pulmonary nodule imaging

Sensitivity of AI lung nodule detection

Identifying a gold standard for the diagnosis of the benign or malignant nature of pulmonary nodules with AI remains a critical issue (14). The accuracy, cost, patient willingness, privacy and security of AI diagnosis are all huge challenges. The application of AI in the detection of pulmonary nodules has been in the process of development, evolving from expert systems to classical machine learning techniques, and more recently, to DL approaches; among them, the early CAD “expert system” based on X-rays had a limited level of intelligence and mainly relied on flowcharts, knowledge bases and statistical methods, etc. to assist clinical decision-making. With the development of image recognition technology, DL-based radiomics technology has gradually emerged in the detection of pulmonary nodules, this is a transformative technical approach that uses AI algorithms (especially machine learning and DL) to extract quantitative features from multimodal medical images, such as CT, positron emission tomography-computed tomography (PET-CT). The technique integrates these features with multi-omics data such as clinical and genomic ones to reveal disease characteristics invisible to the human eye, ultimately achieving precise disease diagnosis, prognosis prediction, and personalized treatment decisions. Currently, the diagnosis of pulmonary nodules by AI is principally based on DL models, such as CNNs, 3D-CNN, Deep Cubical Nodule Transfer Learning Algorithm (DeepCUBIT) model, recurrent neural networks (RNNs), generative adversarial networks (GANs), and transformers. Through the use of lung imaging data, deep neural networks can continuously learn and approximate complex real-world models and thus possess a strong learning ability and flexible diagnostic function (15). Among various algorithms, the CNN algorithm is the majority (16). The main principle of this program is to repeatedly perform automatic feature extraction on images, continuously optimize and classify them to achieve automatic recognition of pulmonary nodules, and make full use of the spatial context information of 3D pulmonary nodules. The use of a multi-view strategy can enhance the classification and sensitivity of two-dimension convolutional neural network (2D CNN) for pulmonary nodules. At the same time, it reduces the false positive rate and has the characteristics of high sensitivity and strong robustness. In their study, based on the different proportions of solid components and ground-glass components in pulmonary nodules on lung ct images, Cai et al. (17) classify nodules into three types: solid nodules (SNs), partially solid nodules (PSNs) and ground-glass nodules (GGNs). The positive diagnostic rates of simple CT and CAD are compared. The results showed that the positive rates of CAD were 95.00%, 96.67%, and 96.00% respectively, all higher than those of simple CT examination (positive rates were 62.50%, 66.67%, and 60.00%). Li et al. (18) compared the malignancy detection rate between the AI and human interpretation of images in 245 nodules from 82 patients with pulmonary nodules. They found that the malignancy detection rate of human interpreters was 30.20% (74/245), while that of AI was 41.22% (101/245). Compared to human interpretation, the specificity of AI film reading for the malignant diagnosis of pulmonary nodules decreased, while the sensitivity and negative predictive value increased. The differences were statistically significant (all P<0.05). This was due to the ability of the AI’s imaging assistance diagnosis system, which could slice and rotate images, enabling rapid detection of pulmonary nodules within seconds. Additionally, the AI system had a higher detection rate for small lesions, providing easier identification of nodules with diameters of 5 to 10 mm and those smaller than 5 mm, thus reducing missed detection (19). Meanwhile, Ye et al. (20) compared three modalities in their ability to detect pulmonary nodules based on CT images: human diagnosis alone, AI diagnosis alone, and CAD. They discovered that CAD could help reduce the reading time and effectively improve the diagnostic sensitivity for pulmonary nodules. Furthermore, Zhang et al. reported that the AI-assisted imaging diagnosis based on AI iterative algorithm [AI-Infused Development (AIID)] could significantly improve the quality of low-dose CT images while maintaining the sensitivity of AI in diagnosing pulmonary nodules. The CAD film reading time is shortened by 63.3% compared with manual film reading, reducing false positives and image noise (21).

Specificity of AI for lung nodule detection

Although the high detection rates that AI can achieve in detecting pulmonary nodules are worth noting, it is also crucial not to overlook its specificity. A study (22) has found that normal lung tissue also shows nodular manifestations on the central area of CT images, especially near the hilum. It is difficult for doctors to distinguish pulmonary nodules from vascular and sputum embolism. The most common reasons for missed diagnosis are their small volume (<7 mm) or the unclear CT images when the pulmonary nodules are located in the central area of the hilum. According to a study by Torres et al., the false negatives of small pulmonary nodules (especially those less than 6 mm and subsolid nodules) in a single examination by radiologists is 35%, and the accurate detection of pulmonary nodules remains a clinical challenge (23). However, the research by Peters et al. demonstrated that with the assistance of CAD, doctors improved the average detection rate of pulmonary nodules (77%, P<0.001) and the accuracy of segmental localization (68%, P<0.001) (24). Lan et al. designed a DL network-based model to assist physicians in the imaging-based diagnosis of pulmonary nodules. The sensitivity for small nodules was significantly improved, while the overall false positives were reduced and sensitivity was increased when physicians diagnosed pulmonary nodules (25). In their study, Liu et al. (26) used adaptive iterative dose reduction Adaptive Iterative Dose Reduction 3D (AIDR 3D) technology to reduce the noise of CT images and optimize the images. The combined Dr. Wise pulmonary nodule auxiliary diagnosis system was applied to the chest simulation model experiment. The research results revealed that the positive rate was the highest at 82.2% in the low-dose mode. At the same time, it was found that the true positive rate of hilar nodules (66.67%) was significantly lower than that of subpleural and pulmonary parenchyma (88.89%, 84.44%), which might be related to the relatively high noise in the mediastinal area of the chest. The positive rate of pulmonary nodules with a diameter of ≤5 mm was only 59.26%, which was much lower than that of nodules with a diameter of ≥8 mm (94.83%). In conclusion, for various AI models, there are differences in the detection effects on pulmonary nodules of different densities, locations, and sizes. The smaller the volume of pulmonary nodules, the lower the detection sensitivity and detection rate of AI. For the location of pulmonary nodules, those that are not connected to blood vessels, bronchi or pleura have a higher detection rate.

As fixed programs, AI algorithms may have inherent flaws that can cause the misjudgment and misdiagnosis of nodule-like imaging features in certain cases, thereby resulting in false positives in diagnosis. Therefore, current hotspots in AI lung cancer research include achieving cross-scale intelligence; generating a unified modeling of imaging, molecular, and clinical data; and improving the specificity of models for pulmonary nodules while reducing the false-positive rate. Overall, with the assistance of AI, the consistency between the detection of pulmonary nodules and pathological results can be greatly improved. In one study (27), CT images were analyzed using the 3D DenseNet model, and postoperative pathological results were used as the gold standard for differentiation. The results showed that the AUC of the AI group and the doctor group was 0.91 and 0.767 respectively, the sensitivities were 94.3% and 80.5% respectively, and the specificities were 70.2% and 63.3% respectively. AI models have a high performance in differentiating benign and malignant pulmonary nodules. In addition, the results also showed that it also had a relatively high efficiency in predicting the invasion degree of lung adenocarcinoma (AUC was 0.808, sensitivity was 71.0%, and specificity was 89.7%). Wang et al. used a center-focused CNN for the segmentation of pulmonary nodules, fully utilizing various features of pulmonary nodules. The results showed excellent performance in differentiating pulmonary nodules in the pleura (28). Wang et al. developed a CAD system-based technology called Central Focused Convolutional Neural Network (CF-CNN), which combines the automatic lung segmentation method with nodule detection using dynamic programming technology. The Dice similarity coefficient (DSC) is a widely used metric at present, serving as an evaluation index to measure the overlap degree between two segmentation results. In the study, 967 pulmonary nodules were analyzed from public LIDC datasets including Guangdong General Hospital (GDGH) and independent datasets. The results show that the DSC of LIDC is 82.15% and that of GDGH is 80.02%, both of which are superior to traditional methods such as Level Set (60.63%), Graph Cut (68.90%), and U-Net (79.50%). This technology has achieved high-precision segmentation of pulmonary nodules in complex CT images, providing an automated tool for the early diagnosis of lung cancer (29). Meanwhile, Yuan et al. developed a multipooling 3D CNN [Multi-Pooling 3D Convolutional Neural Network (MP-3D-CNN)] detection model by using multipathway extraction and the fusion of 3D feature information related to pulmonary nodules. The research team designed three independent information input paths (Archi-1/2/3), respectively using three different-sized receptive fields to extract local detail, medium-range and global context features, and integrated multi-scale information through feature map splicing to enhance the recognition ability of nodules of different sizes and shapes. This allowed the model to quickly and automatically adapt to changes in the shape, size and others information of pulmonary nodules, thus more effectively reducing false positives in pulmonary nodule detection. The results verify that MP-3D-CNN demonstrated a high diagnostic specificity of 0.998. The multi-image output pathway of the 3D-CNN model can obtain more spatial information of pulmonary nodules as compared to the 2D-CNN model, which greatly improves the diagnosis of pulmonary nodules (30). Zhang et al. based on a multi-scale attention (MSA) block that can fully utilize multiscale image information. They further constructed a candidate nodule detection network for 3D Faster R-CNN and combined with a U-Net-based encoder-decoder structure, as well as a false positive reduction network based on the 3D-CNN with MSA blocks. They trained the model on a large amount of CT image data from the LUNA16 and Tianchi datasets. Their results indicated that the proposed 3D detection system integrating multiscale features and attention mechanisms demonstrated significantly improved sensitivity and false positive control [achieving a candidate per milliliter (CPM) of 0.927 on LUNA16] but missed detecting small-sized nodules (<5 mm) (CPM =0.713) (31). The CPM score is a performance index that comprehensively considers sensitivity and false positive control ability, calculated as the average sensitivity under an average false positive number of 0.125, 0.25, 0.5, 1, 2, 4, and 8 per scan. The CPM of the research team’s AI model is as high as 0.927, demonstrating its advanced nature and practicality in the task of pulmonary nodule detection.

The CAD system for pulmonary nodules is evolving from single image analysis to multi-modal intelligent diagnosis, with the integration of three-dimensional DL, radiomics and clinical data becoming the core driving force. It enhances the early diagnosis and treatment outcomes of diseases such as lung cancer, thereby benefiting the prognosis of patients (32). Moreover, CAD can provide important auxiliary reference value for clinical diagnosis in differentiating benign and malignant pulmonary nodules. However, pulmonary nodules may be stereoscopic, multidimensional, or variable in nature, and thus, at this time, AI cannot completely replace the clinical diagnosis pulmonary nodules performed by radiologists. In the process of AI-based diagnosis, a detailed analysis of the size, location, density, and clinical data of the nodules can clarify the nature of the nodules, thereby reducing the rate of misdiagnosis (33). Saqi et al. (34) combined Entrop-based Improved Tree Learning Random Forest (EITL-RF) and CNN multi-class segmentation AI models. The tumor microenvironment (TME) of fibrotic and non-fibrotic lung non-small cell lung cancer (NSCLC) was evaluated using this comprehensive model. The results verify that the overall classification accuracy is 94%, the sensitivity is 90%, and the F1 score is 91%, providing an objective and multi-dimensional evaluation for lung cancer. In conclusion, AI has made significant progress in the detection, segmentation, and diagnosis of benign and malignant pulmonary nodules based on CT image data. In the future, continuous efforts are needed in establishing multimodal models, expanding the detection range, predicting the prognosis of patients, building databases, establishing real-time dynamic learning systems, and constructing ethical frameworks.

AI detection of pulmonary nodule metastasis

TNM staging remains critical for personalized treatment among patients with lung cancer and can even affect patient prognosis and survival. Within this staging paradigm, the status of lymph node metastasis (LNM) in lung cancer is crucial. Currently, lobectomy combined with lymph node dissection is the main surgical approach for early-stage lung cancer. However, the preoperative diagnosis of LNM by clinical physicians and radiologists based on CT images is often subjective, relies solely on the clinicians’ experience, and is thus susceptible to error. Recently developed DL-based methods can significantly reduce the difficulty of pulmonary nodule feature extraction and possess certain clinical application value (35). Zhao et al. built a model based on 3D Multi-scale, Multi-task, and Multi-label classification network (3M-CN) to predict LNM according to the characteristics of pulmonary nodules and the clinical factors of LNM, which has broad clinical application prospects. The research results demonstrated that the AUC of 3M-CN in the internal and external test datasets reached 0.945 and 0.948 respectively, which was significantly superior to the baseline model (0.901/0.812) and the radiomics method (0.885/0.801). At the same time, the average AUC increased by 3.5% (internal) and 5.3% (external), and the network improves the performance of multi-label classification (36). Zuo et al. designed a nomogram model to predict the preoperative extent of visceral pleural invasion (VPI) in early-stage lung adenocarcinoma by analyzing the CT images and other clinical manifestations of pulmonary nodules. The AUC of this model in the training queue and the validation queue is 0.890 and 0.864, respectively. They also found that features such as the longitudinal diameter and skewness of pulmonary nodules were associated with VPI and that the model demonstrated good predictive value and application prospects (37). This nomogram provides a quantitative tool for preoperative assessment of VPI and is helpful in guiding surgical decisions (such as choosing lobectomy over sublobectomy), but its wide application requires further multi-center validation. The Sato research team used the commercially available CAD system, integrating with SYNAPSE SAI Viewer V2.4, to measure the differences in tumor size of lung adenocarcinoma on CT images, and evaluate its effect on LNM. The research results confirmed that CAD system (0.82–0.85) was stronger than physician reading in predicting LNM (AUC: 0.75–0.80). Therefore, AI-based CAD systems can predict LNM and prognosis with complete repeatability and are not affected by image display conditions. These results may provide a highly reproducible method for identifying T descriptors in lung adenocarcinoma and may contribute to treatment decisions and efficacy (38).

The above-mentioned models have demonstrated the positive prediction of pulmonary nodule metastasis by AI models, which is in line with the clinical diagnostic process and can be extended to the intelligent auxiliary diagnosis of other diseases in the future. Despite those promising research, few prospective studies on AI models regarding the metastasis of lung cancer to surrounding tissues, especially LNM, have been conducted, and the focus has largely been on early-stage lung cancer. Therefore, it is necessary to conduct more in-depth research to ensure AI can predict LNM across various stages and grades of lung cancer, thereby enhancing clinical applications and medical outcomes.

Applications of AI in pulmonary nodule pathology

NSCLC cancer and small cell lung cancer constitute the two main types of lung cancer, with NSCLC accounting for approximately 80–85% of all lung cancer cases (39). According to histopathology, lung cancer can also be divided into subtypes, such as squamous cell carcinoma and adenocarcinoma; in turn, according to the growth mode and morphology, lung adenocarcinoma can be divided into in situ adenocarcinoma, minimally invasive adenocarcinoma, acinar adenocarcinoma, adherent growth adenocarcinoma, solid adenocarcinoma, papillary adenocarcinoma, micropapillary adenocarcinoma, and other classifications. According to the details of this classification, individualized treatment plans can be devised (40). Fine-needle aspiration biopsy or postoperative pathological examination is commonly used in clinical practice to determine the pathological type of pulmonary nodules, but as invasive procedures, they involve certain risks and may increase the likelihood of cancer metastasis, imposing unnecessary burdens on the patient (41). Therefore, non-invasive operations like AI technology have certain significance in clinical practice. Research indicates that extracting CT image texture and analyzing its basic imaging features can improve the accuracy of preoperatively predicting lung adenocarcinoma (42). Teramoto et al. (43) developed an automatic pathological classification model for lung cancer cells observed in microscopic images using a deep CNN or Deep Convolutional Neural Network (DCNN) and applied GANs to enhance the DCNN. After examining 298 tissue pathology images, the model classified adenocarcinoma, squamous cell carcinoma, and small cell lung cancer with accuracy rates of 89%, 60%, and 70%, respectively. Xu et al. (44) demonstrated that AI can assess the pathological classification and degree of infiltration of GGNs, and through AI recognition technology based on DL models, the pathological classification of GGN can be accurately distinguished preoperatively. In their study, preoperative AI assessment had a higher accuracy rate than did postoperative pathological results in terms of pathological classification and infiltration degree of GGNs, significantly improving clinicians’ work efficiency and enhancing the ability of junior physicians to diagnose GGNs. Wang et al. (45) proposed an interpretable multitask attention learning network (IMAL-Net) model, which extracts radiomics features related nodule size, shape, and texture and uses a feature selection mechanism to reveal the quantitative relationship between important radiomics features and pulmonary nodules. They found the model had excellent performance in predicting pulmonary nodules and can be used for the early screening of invasive lung cancer, with high clinical application value. Additionally, Coudray et al. (46) devised a DCNN and trained it using slice images obtained from The Cancer Genome Atlas. The results showed that their network model could accurately and automatically classify lung tissue from pathology images into adenocarcinoma, squamous cell carcinoma, and normal lung tissue. Khosravi et al. (47) constructed a novel fine-tuned pretrained CNN model capable of distinguishing between squamous cell carcinoma and adenocarcinoma in high-resolution magnified images with an accuracy of up to 75–90%. Overall, the development of AI-based pathological diagnosis systems for lung cancer has improved the prediction of pathological results of lung nodules, improved the efficiency of pathologists and provided a new path for progress in pathology.

Evidence shows that multi-module AI models, when combined with methods such as pathology and liquid biopsy, can be optimized to assist in early diagnosis and risk prediction, and improve diagnostic accuracy(48). Qiao et al. (49) used the AI-assisted system named SANMED for the segmentation and 3D reconstruction of lung nodules and combined it with circulating genetically abnormal cell (CAC) detection to form a multimodule AI model. This model demonstrated considerable ability in predicting the pathological subtypes of subsolid-type lung adenocarcinoma, suggesting that integrating AI with liquid biopsy can enhance the accuracy of predicting invasive lung adenocarcinoma. In their study, Ye et al. found that radiomics and deep transfer learning (Rad-DTL) features are potential biomarkers for the preoperative CT prediction of high-risk pathological lung nodules and developed and validated a DL model for preoperative CT prediction of high-risk pathological lung nodules by integrating radiomics features, Rad-DTL features, and clinical features (50).

However, at this stage, the use of AI for pathological diagnosis remains in the research phase, and pathological examination will remain the gold standard in the detection of histological subtypes of lung cancer in the foreseeable future. This is because standardized processes for specimen processing, section staining, and lesion image annotation in pathological sections have not yet been established, and this has resulted in insufficient pathological data available for AI training and learning, which reduces the reliability of diagnosis. All datasets come from a single population, potentially overestimating performance. Research on the sensitivity of AI to rare subtypes such as micropapilla/solid subtypes is scarce. The opacity of AI decision-making logic may lead to the results obtained by the model not necessarily convincing doctors. The World Health Organization (WHO) revised the pathological classification of lung cancer in 2021. The precursor glandular lesions (PGL) were removed from lung adenocarcinoma and classified as benign lesions. Therefore, the CAD model needs to be optimized along with the update of pathological classification to evaluate the PGL (51). For example, in Huang et al.’s study, they found that the classification change of PGL significantly affected the performance of AI (the accuracy rate of benign diagnosis decreased from 72.86% to 43.48%) (52). AI is driving a transformation in pathological examination by enhancing diagnostic consistency, assisting with tedious tasks (such as cell counting), and predicting clinical outcomes. However, issues regarding model interpretability, validation, and regulatory concerns need to be addressed. In addition, liquid biopsy faces challenges in clinical implementation, including standardization of pre-analysis variables and data sharing, but AI can significantly enhance its performance (48). Future directions of research should include multimodal integration and human-machine collaborative processes (53).

Applications of AI in lung nodule genomics

Radiogenomics has long been a field of intense interest in lung cancer research and has been applied in patient prognosis and personalized treatment and disease management. Lung tumors of the same pathological subtype may differ significantly at the molecular level, thus producing different responses to treatment. Genetic alterations with prognostic and predictive significance in NSCLC include epidermal growth factor receptor (EGFR) mutations and anaplastic lymphoma kinase (ALK) rearrangements, among others. EGFR mutation status is key to selecting appropriate targeted therapy in clinical treatment, with EGFR tyrosine kinase inhibitors being first-line targeted drugs that can effectively alleviate patients’ conditions and prolong survival (54). However, invasive procedures such as needle biopsy or surgical sampling are often used to collect lung nodule samples in clinical practice, which is not suitable for all patients with lung cancer. Therefore, there has been increased interest in applying AI DL models based on radiomics features from CT and PET as potential biomarkers for the noninvasive tumor characterization in the baseline staging of lung cancer and response assessment. This process involves extracting digital features of specific biomarkers from images to evaluate the correlation between imaging features and genomic maps. As an auxiliary tool, AI can accelerate the initial assessment and treatment decision-making for patients with NSCLC, and more prospective studies are needed to standardize DL algorithms and integrate multimodal imaging with clinical data to improve model accuracy (55). A study on radiogenomics in NSCLC indicate that CT imaging radiomics features can predict EGFR and/or KRAS mutation status, which can serve as a feasible method for decoding tumor heterogeneity, thereby facilitating the individualized treatment of patients (56). Moulson et al. (57) analyzed real-world treatment patterns and clinical outcomes in patients with late-stage EGFR-mutated NSCLC by using a large language model (LLM) system named DARWEN^TM. The results indicated that the model had high application value in extracting data, predicting the efficacy of targeted drug therapy, and assisting in the development of novel drugs. Nguyen et al. (58) reported that an AI model showed moderate diagnostic efficacy in predicting EGFR mutation status in patients with NSCLC, with an AUC of 0.789 and an overall sensitivity and specificity of 72.2% and 73.3%, respectively. Moreover, Zhou et al. (59) developed an imaging biomarker model based on dual-energy CT venous phase images to predict EGFR mutations in lung adenocarcinoma, which demonstrated excellent predictive ability and high reference value for the clinical treatment of lung adenocarcinoma. Coudray et al. speculated that some gene mutations can change the arrangement of lung cancer cells in slice images and predicted the most commonly mutated genes in adenocarcinoma by training a neural network. They found that certain genes, including serine-threonine kinase 11 (STK11), EGFR, FAT atypical cadherin 1 (FAT1), SET binding protein 1 (SETBP1), Kirsten rat sarcoma virus (KRAS), and tumor protein p53 (TP53), could be predicted through pathological images at an accuracy of 73.3–85.6% (46). Liu et al. (60) devised an immune scoring system (patho-immunoscore) based on the Phikon model, which could effectively predict the efficacy of immunochemotherapy in patients with advanced non-squamous NSCLC. Gui et al. (61) proposed a new multi-task learning method, AIR-Net, based on the shared encoder ResNet-50 for predicting the EGFR mutation status on CT images, which achieved an AUC of approximately 0.86 in testing, indicating broad application. In addition, Shiri et al. (62) introduced ComBat coordination into radiomics models, significantly improving the models’ image-based predictive performance and effectively enhancing the efficiency of pulmonary nodule diagnosis. Tang et al. (63) constructed a predictive model by applying CT imaging radiomics features and analyzed the correlation between the radiomics features of patients with NSCLC and EGFR gene mutation status. They obtained five relatively important imaging features that reflect the strong correlation between the genetic phenotype of NSCLC and the normal lung tissue surrounding the lung cancer. This provides a degree of support for the development and implementation of individualized treatment plans for patients with NSCLC. Batra et al. (64) developed the AI-Based Predictive System (AIPS) model for non-invasive prediction of EGFR genotypes and further trained the AIPS-Nodule (AIPS-N) model for the automatic detection and characterization of pulmonary nodules, and explored the value of combining its results with the clinical factors in the AIPS-Mutation (AIPS-M) model in predicting EGFR genotypes. The results showed that after combining CT features, the AUC increased from 0.60 to 0.90, effectively distinguishing the wild type and mutant type of EGFR and assisting in the stratification of targeted therapy, and can provide personalized plans for clinical care, diagnosis, and treatment. Trials on AI-mediated proteomics and multiple biomarker panels are being conducted to achieve more effective detection of various types of lung cancer (65).

Many of the above models can predict mutations such as EGFR/ALK. Whether they can reduce unnecessary biopsies in clinical practice requires further clinical verification. Evidence shows that the EGFR mutation rate in non-smoking adenocarcinoma in Asia is as high as 50%, while in North America, it is only 15% (66). The accuracy of different models in different races or countries is unknown. Further prospective research and the design of more advanced DL models are still needed. In the future, the synergistic combination of AI and tissue biopsy may provide an attractive diagnostic alternative to traditional genomic analysis as a potential tool for personalized lung cancer management.

Future

With the widespread clinical application of high-resolution low-dose spiral CT and the increased public awareness of self-health, a larger number of pulmonary GNNs are being discovered in clinical practice (67). As a result, misdiagnosis and missed diagnosis of small nodules and other conditions are increasing, and the pressure faced by radiologists is growing in kind (68). The potential applications of AI in the healthcare industry are continuously being explored and have been accompanied by breakthroughs in key technologies such as neural networks and DL algorithms. Through continuous systematic learning, the detection and evaluation of pulmonary nodules have become increasingly accurate. Deeply analyzing CT image data and quantitatively analyzing tumor heterogeneity can reveal imaging features and signs that the human eye cannot detect. This can assist physicians in quickly identifying pulmonary nodules and differentiating their benignity or malignancy, which is particularly helpful for improving the work efficiency of physicians with less experience (69,70). AI plays a crucial role in disease diagnosis and monitoring, medical efficacy evaluation, survival prediction, drug trials, and health management (71). However, AI DL models in medical systems are still in the development stage and are limited in handling complex cases and multimodal fusion in lung cancer diagnosis and treatment (72). The diversity of the research population in terms of ethnicity, gender, and age can lead to biases in the accuracy of AI DL models in diagnosing the benignity and malignancy of pulmonary nodules (8). Furthermore, the prediction of pulmonary nodules by CAD systems affects clinicians’ decisions and the choice of diagnostic and treatment plans for patients, potentially leading to a lack of trust. Therefore, there remains considerable work to be done in refining the application of AI DL models in clinical practice (73,74).

In summary, there have been significant advancements in AI technology for the early screening of pulmonary nodules, the differentiation of benign and malignant lesions, pathological diagnosis and classification, and radiogenomics. Based on imaging and related disciplines, AI models have become increasingly popular in pulmonary nodule diagnosis, assisting clinicians in disease diagnosis and differentiation. However, the further application AI in the field of lung cancer requires the establishment of a pulmonary nodule database and the formation of a unified standard for pulmonary nodule diagnosis. Compared to traditional diagnostic and treatment methods, AI models possess greater research potential in the diagnosis and prognosis of lung cancer and can better inform the precision of treatment of patients with lung cancer.

Conclusions

Therefore, with the advancement of computer technology, AI will play an increasingly important role in various fields. AI enables more precise screening, diagnosis, and personalized treatment in lung cancer diagnosis. Although AI has been proven to be a valuable tool, it will not replace doctors but rather complement their professional knowledge. The key to the future development of medicine lies in the collaboration between AI and doctors, with patients at the center, to provide better medical conditions and promote the progress of lung cancer prevention and treatment.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1512/rc

Peer Review File: Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1512/prf

Funding: This work was supported by the Jiangsu Provincial Health Commission Scientific Research Project (No. M2024101).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2025-1512/coif). All authors report that this work was supported by the Jiangsu Provincial Health Commission Scientific Research Project (No. M2024101). The authors have no other conflicts of interest to declare.

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.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bade BC, Dela Cruz CS. Lung Cancer 2020: Epidemiology, Etiology, and Prevention. Clin Chest Med 2020;41:1-24. [Crossref] [PubMed]
2. Dziedzic R, Marjański T, Rzyman W. A narrative review of invasive diagnostics and treatment of early lung cancer. Transl Lung Cancer Res 2021;10:1110-23. [Crossref] [PubMed]
3. National Lung Screening Trial Research Team. Reduced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med 2011;365:395-409. [Crossref] [PubMed]
4. Rossi A, Maione P, Colantuoni G, et al. Screening for lung cancer: New horizons? Crit Rev Oncol Hematol 2005;56:311-20. [Crossref] [PubMed]
5. Jungblut L, Blüthgen C, Polacin M, et al. First Performance Evaluation of an Artificial Intelligence-Based Computer-Aided Detection System for Pulmonary Nodule Evaluation in Dual-Source Photon-Counting Detector CT at Different Low-Dose Levels. Invest Radiol 2022;57:108-14. [Crossref] [PubMed]
6. Liu YB, Li Q, Zhou W, et al. Chest CT Combined with AI Diagnosis System in the Diagnosis of Patients with Suspected Pulmonary Nodules and the Evaluation Value of Nodule Types. Progress in Modern Biomedicine 2022;22:955-959+949.
7. Cao W, Wu R, Cao G, et al. A comprehensive review of computer-aided diagnosis of pulmonary nodules based on computed tomography scans. IEEE Access 2020;8:154007-23.
8. Hamidian S, Sahiner B, Petrick N, et al. 3D Convolutional Neural Network for Automatic Detection of Lung Nodules in Chest CT. Proc SPIE Int Soc Opt Eng 2017;10134:1013409. [Crossref] [PubMed]
9. Sha L, Osinski BL, Ho IY, et al. Multi-Field-of-View Deep Learning Model Predicts Nonsmall Cell Lung Cancer Programmed Death-Ligand 1 Status from Whole-Slide Hematoxylin and Eosin Images. J Pathol Inform 2019;10:24. [Crossref] [PubMed]
10. Chen J, Ming M, Huang S, et al. AI-enhanced diagnostic model for pulmonary nodule classification. Front Oncol 2024;14:1417753. [Crossref] [PubMed]
11. Yang D, Miao Y, Liu C, et al. Advances in artificial intelligence applications in the field of lung cancer. Front Oncol 2024;14:1449068. [Crossref] [PubMed]
12. Yoon SH, Park S, Jang S, et al. Use of artificial intelligence in triaging of chest radiographs to reduce radiologists' workload. Eur Radiol 2024;34:1094-103. [Crossref] [PubMed]
13. Ost DE. Artificial intelligence applications for the diagnosis of pulmonary nodules. Curr Opin Pulm Med 2025;31:344-51. [Crossref] [PubMed]
14. Cui S, Ming S, Lin Y, et al. Development and clinical application of deep learning model for lung nodules screening on CT images. Sci Rep 2020;10:13657. [Crossref] [PubMed]
15. Huang S, Yang J, Shen N, et al. Artificial intelligence in lung cancer diagnosis and prognosis: Current application and future perspective. Semin Cancer Biol 2023;89:30-7. [Crossref] [PubMed]
16. Valente IR, Cortez PC, Neto EC, et al. Automatic 3D pulmonary nodule detection in CT images: A survey. Comput Methods Programs Biomed 2016;124:91-107. [Crossref] [PubMed]
17. Cai SH, Lin QJ, Yang SM, et al. Clinical value of CT and AI pulmonary nodule diagnosis system for diagnosing pulmonary nodules and differentiating subtypes. Medical Equipment 2024;30-3.
18. Li B, Liu YB, Cui Y, et al. Value Analysis of Image Reading in Artificial Intelligence-Assisted Diagnostic System for Differentiating the Nature of Lung Nodules. Modern Medicine and Health Research 2024;110-3.
19. Li T, Li XD, Liu JY. Clinical Value of Artificial Intelligence in the Diagnosis of Pulmonary Nodules. Chinese General Practice 2020;23:828-31.
20. Ye WW, Liu BH, Guo TC. Clinical application of deep learning artificial intelligence in qualitative diagnosis of pulmonary nodules. Journal of Imaging Research and Medical Applications 2024;(3):8-10+16.
21. Zhang BP, Li A, Li YH, et al. Impacts of artificial intelligence iterative reconstruction algorithm on the image quality and computer-aided pulmonary nodule detection at ultra-low dose chest. Journal of Zunyi Medical University 2024;47:384-91.
22. Del Ciello A, Franchi P, Contegiacomo A, et al. Missed lung cancer: when, where, and why? Diagn Interv Radiol 2017;23:118-26. [Crossref] [PubMed]
23. Torres EL, Fiorina E, Pennazio F, et al. Large scale validation of the M5L lung CAD on heterogeneous CT datasets. Med Phys 2015;42:1477-89. [Crossref] [PubMed]
24. Peters AA, Wiescholek N, Müller M, et al. Impact of artificial intelligence assistance on pulmonary nodule detection and localization in chest CT: a comparative study among radiologists of varying experience levels. Sci Rep 2024;14:22447. [Crossref] [PubMed]
25. Lan CC, Hsieh MS, Hsiao JK, et al. Deep Learning-based Artificial Intelligence Improves Accuracy of Error-prone Lung Nodules. Int J Med Sci 2022;19:490-8. [Crossref] [PubMed]
26. Liu Q, Zeng YM, Sun JK, et al. Analysis of Influencing Factors on Pulmonary Nodule Detection by Computed Tomography with Artificial Intelligence: A Phantom Study. Computerized Tomography Theory and Applications 2024;33:471-7.
27. He Z, Song W, Deng Z, et al. A preliminary clinical investigation on artificial intelligence imaging system to predict benign, malignant, and infiltrative lung nodules. Journal of Clinical Pulmonary Medicine 2023;28:1783-1787+1792.
28. Wang S, Zhou M, Liu Z, et al. Central focused convolutional neural networks: Developing a data-driven model for lung nodule segmentation. Med Image Anal 2017;40:172-83. [Crossref] [PubMed]
29. Wang J, Dobbins JT 3rd, Li Q. Automated lung segmentation in digital chest tomosynthesis. Med Phys 2012;39:732-41. [Crossref] [PubMed]
30. Yuan H, Fan Z, Wu Y, et al. An efficient multi-path 3D convolutional neural network for false-positive reduction of pulmonary nodule detection. Int J Comput Assist Radiol Surg 2021;16:2269-77. [Crossref] [PubMed]
31. Zhang H, Peng Y, Guo Y. Pulmonary nodules detection based on multi-scale attention networks. Sci Rep 2022;12:1466. [Crossref] [PubMed]
32. Oliveira SP, Neto PC, Fraga J, et al. CAD systems for colorectal cancer from WSI are still not ready for clinical acceptance. Sci Rep 2021;11:14358. [Crossref] [PubMed]
33. Su YC, Zhang XQ. Artificial Intelligence-assisted Diagnosis in Detecting Lung Nodules and Differentiating Benign from Malignant Nodules. Computerized Tomography Theory and Applications 2024;33:325-31.
34. Saqi A, Liu Y, Politis M G, et al. Combined expert-in-the-loop—random forest multiclass segmentation U-net based artificial intelligence model: evaluation of non-small cell lung cancer in fibrotic and non-fibrotic microenvironments. Journal of Translational Medicine 2024;22:640. [Crossref] [PubMed]
35. Zhao X, Wang X, Xia W, et al. A cross-modal 3D deep learning for accurate lymph node metastasis prediction in clinical stage T1 lung adenocarcinoma. Lung Cancer 2020;145:10-7. [Crossref] [PubMed]
36. Zhao X, Wang X, Xia W, et al. 3D multi-scale, multi-task, and multi-label deep learning for prediction of lymph node metastasis in T1 lung adenocarcinoma patients' CT images. Comput Med Imaging Graph 2021;93:101987. [Crossref] [PubMed]
37. Zuo Z, Li Y, Peng K, et al. CT texture analysis-based nomogram for the preoperative prediction of visceral pleural invasion in cT1N0M0 lung adenocarcinoma: an external validation cohort study. Clin Radiol 2022;77:e215-21. [Crossref] [PubMed]
38. Sato J, Yanagawa M, Nishigaki D, et al. Radiologists Versus AI-Based Software: Predicting Lymph Node Metastasis and Prognosis in Lung Adenocarcinoma From CT Under Various Image Display Conditions. Clin Lung Cancer 2025;26:58-71. [Crossref] [PubMed]
39. Global Burden of Disease Cancer Collaboration. Global, Regional, and National Cancer Incidence, Mortality, Years of Life Lost, Years Lived With Disability, and Disability-Adjusted Life-Years for 29 Cancer Groups, 1990 to 2017: A Systematic Analysis for the Global Burden of Disease Study. JAMA Oncol 2019;5:1749-68. [Crossref] [PubMed]
40. Hung JJ, Yeh YC, Jeng WJ, et al. Predictive value of the international association for the study of lung cancer/American Thoracic Society/European Respiratory Society classification of lung adenocarcinoma in tumor recurrence and patient survival. J Clin Oncol 2014;32:2357-64. [Crossref] [PubMed]
41. Lu L, Yong XM, Yu TF. Comparative Analysis of CT And Pathological Findings of Lung Adenocarcinoma with Mixed Ground-Glass Nodule. Chinese Journal of CT and MRI 2022;38-40.
42. Jin ZF, Chen XM, Feng B, et al. CT texture features in differentiation of minimally invasive andinvasive adenocarcinoma manifesting assubsolid pulmonary nodules. Chinese Journal of Medical Imaging Technology 2019;35:691-5.
43. Teramoto A, Tsukamoto T, Kiriyama Y, et al. Automated Classification of Lung Cancer Types from Cytological Images Using Deep Convolutional Neural Networks. Biomed Res Int 2017;2017:4067832. [Crossref] [PubMed]
44. Xu GA, Zhu LY, Xu J, et al. Application of Artificial Intelligence Pulmonary Nodule Auxiliary Diagnosis System in Assessment of the Degree of Pathological Infiltration of Pulmonary Ground Glass Nodule. Journal of Nanchang University (Medical Sciences) 2022;53-6.
45. Wang J, Yuan C, Han C, et al. IMAL-Net: Interpretable multi-task attention learning network for invasive lung adenocarcinoma screening in CT images. Med Phys 2021;48:7913-29. [Crossref] [PubMed]
46. Coudray N, Ocampo PS, Sakellaropoulos T, et al. Classification and mutation prediction from non-small cell lung cancer histopathology images using deep learning. Nat Med 2018;24:1559-67. [Crossref] [PubMed]
47. Khosravi P, Kazemi E, Imielinski M, et al. Deep Convolutional Neural Networks Enable Discrimination of Heterogeneous Digital Pathology Images. EBioMedicine 2018;27:317-28. [Crossref] [PubMed]
48. Ignatiadis M, Sledge GW, Jeffrey SS. Liquid biopsy enters the clinic - implementation issues and future challenges. Nat Rev Clin Oncol 2021;18:297-312. [Crossref] [PubMed]
49. Qiao JJ, Zhang GR, Chen HJ, et al. Diagnostic value of artificial intelligence combined with circulating chromosomal abnormal cells in pathological subtypes of subsolid nodular lung adenocarcinoma. Journal of Clinical Pulmonary Medicine 2025;30:661-6.
50. Ye G, Wu G, Li K, et al. Development and Validation of a Deep Learning Radiomics Model to Predict High-Risk Pathologic Pulmonary Nodules Using Preoperative Computed Tomography. Acad Radiol 2024;31:1686-97. [Crossref] [PubMed]
51. Nicholson AG, Tsao MS, Beasley MB, et al. The 2021 WHO Classification of Lung Tumors: Impact of Advances Since 2015. J Thorac Oncol 2022;17:362-87. [Crossref] [PubMed]
52. Huang WJ, Li QQ, Jing WB, et al. Diagnostic Performance of Artificial Intelligence on Benign and Malignant Pulmonary Nodules Under Different WHO Pathological Classifications of Lung Cancer. Journal of Clinical Radiology 2025;44:76-82.
53. Caranfil E, Lami K, Uegami W, et al. Artificial Intelligence and Lung Pathology. Adv Anat Pathol 2024;31:344-51. [Crossref] [PubMed]
54. Larsen JE, Cascone T, Gerber DE, et al. Targeted therapies for lung cancer: clinical experience and novel agents. Cancer J 2011;17:512-27. [Crossref] [PubMed]
55. Asadian S. Non-Small Cell Lung Cancer Genotype Prediction Utilizing Multiparametric Artificial Intelligence Investigations. Acad Radiol 2024;31:684-5. [Crossref] [PubMed]
56. Shiri I, Maleki H, Hajianfar G, et al. Next-Generation Radiogenomics Sequencing for Prediction of EGFR and KRAS Mutation Status in NSCLC Patients Using Multimodal Imaging and Machine Learning Algorithms. Mol Imaging Biol 2020;22:1132-48. [Crossref] [PubMed]
57. Moulson R, Law J, Sacher A, et al. Real-World Outcomes of Patients with Advanced Epidermal Growth Factor Receptor-Mutated Non-Small Cell Lung Cancer in Canada Using Data Extracted by Large Language Model-Based Artificial Intelligence. Curr Oncol 2024;31:1947-60. [Crossref] [PubMed]
58. Nguyen HS, Ho DKN, Nguyen NN, et al. Predicting EGFR Mutation Status in Non-Small Cell Lung Cancer Using Artificial Intelligence: A Systematic Review and Meta-Analysis. Acad Radiol 2024;31:660-83. [Crossref] [PubMed]
59. Zhou JZ, Fu XL, Zou HY, et al. Radiomics Analysis of Venous Phase Dual-Energy CT Imaging for Predicting EGFR Mutation Status in Lung Adenocarcinoma. Journal of Clinical Radiology 2021;40:1516-20.
60. Liu J, Sun D, Xu S, et al. Association of artificial intelligence-based immunoscore with the efficacy of chemoimmunotherapy in patients with advanced non-squamous non-small cell lung cancer: a multicentre retrospective study. Front Immunol 2024;15:1485703. [Crossref] [PubMed]
61. Gui D, Song Q, Song B, et al. AIR-Net: A novel multi-task learning method with auxiliary image reconstruction for predicting EGFR mutation status on CT images of NSCLC patients. Comput Biol Med 2022;141:105157. [Crossref] [PubMed]
62. Shiri I, Amini M, Nazari M, et al. Impact of feature harmonization on radiogenomics analysis: Prediction of EGFR and KRAS mutations from non-small cell lung cancer PET/CT images. Comput Biol Med 2022;142:105230. [Crossref] [PubMed]
63. Tang CC, Chen AQ, Du XM, et al. Predictive Value of CT Radiomics in Epidermal Growth Factor Receptor Mutations in Non-small Cell Lung Cancer. Chinese Journal of CT and MRI 2023;67-70.
64. Batra U, Nathany S, Nath SK, et al. AI-based pipeline for early screening of lung cancer: integrating radiology, clinical, and genomics data. Lancet Reg Health Southeast Asia 2024;24:100352. [Crossref] [PubMed]
65. Gandhi Z, Gurram P, Amgai B, et al. Artificial Intelligence and Lung Cancer: Impact on Improving Patient Outcomes. Cancers (Basel) 2023;15:5236. [Crossref] [PubMed]
66. Dela Cruz CS, Tanoue LT, Matthay RA. Lung cancer: epidemiology, etiology, and prevention. Clin Chest Med 2011;32:605-44. [Crossref] [PubMed]
67. He J, Li N, Chen WQ, et al. China Guideline for the Screening and Early Detection of Lung Cancer (2021, Beijing). China Cancer 2021;30:81-111. [Crossref] [PubMed]
68. Tang ZX, Wang YM, Zhou LY, et al. Artificial intelligence technologies in lung imaging assisted diagnosis: a review. Chinese Journal of Medical Physics 2022;39:655-60.
69. Fan HY, Sun DD, Zhang Q, et al. A comparative study about the detection capability of lung nodules on CT images based on AI-assisted software among primary and senior interns. Chinese Imaging Journal of Integrated Traditional and Western Medicine 2021;19:175-9.
70. van Leeuwen KG, Schalekamp S, Rutten MJCM, et al. Comparison of Commercial AI Software Performance for Radiograph Lung Nodule Detection and Bone Age Prediction. Radiology 2024;310:e230981. [Crossref] [PubMed]
71. Liu K, Li Q, Ma J, et al. Evaluating a Fully Automated Pulmonary Nodule Detection Approach and Its Impact on Radiologist Performance. Radiol Artif Intell 2019;1:e180084. [Crossref] [PubMed]
72. Afshar S, Afshar S, Warden E, et al. Application of Artificial Neural Network in miRNA Biomarker Selection and Precise Diagnosis of Colorectal Cancer. Iran Biomed J 2019;23:175-83. [Crossref] [PubMed]
73. Chassagnon G, Dohan A. Artificial intelligence: from challenges to clinical implementation. Diagn Interv Imaging 2020;101:763-4. [Crossref] [PubMed]
74. Alexander R, Waite S, Bruno MA, et al. Mandating Limits on Workload, Duty, and Speed in Radiology. Radiology 2022;304:274-82. [Crossref] [PubMed]