Lung cancer remains the leading cause of cancer-related death worldwide. According to GLOBOCAN 2020 data, there were over 2.2 million new cases and approximately 1.8 million deaths globally that year. Because early-stage lung cancer rarely produces symptoms, the majority of patients are diagnosed at an advanced stage or with distant metastases, resulting in a five-year survival rate of only 15-16%. This stark reality makes early detection and timely intervention critical for improving outcomes.
Low-dose CT screening: Data from the National Lung Screening Trial (NLST) established that regular low-dose CT screening of high-risk individuals detects more early-stage cancers and reduces lung cancer mortality. Guidelines from the Fleischner Society (based on mean nodule diameter and type) and the British Thoracic Society (recommending volumetric measurement over 2D) provide frameworks for managing incidentally detected nodules. However, radiologists' sensitivity in detecting lung nodules varies significantly depending on size, shape, location, density, and relationship with adjacent structures.
Scope of the review: This paper from Frontiers in Oncology (2024, published by Yang et al. from Hebei University) provides a comprehensive overview of AI applications across the entire lung cancer clinical pathway: lung nodule detection, pathological classification, gene mutation prediction, treatment strategies (including 3D reconstruction for surgical planning), immunotherapy response prediction, and prognosis. The review also covers core AI/ML/DL algorithms and addresses limitations and future directions.
The authors define the hierarchy of AI techniques relevant to their discussion. Machine learning (ML) encompasses supervised learning, unsupervised learning, and reinforcement learning, with common algorithms including support vector machines (SVM), random forests, decision trees, logistic regression, k-nearest neighbors, and Bayesian networks. Deep learning (DL), a subfield of ML, uses multi-layer artificial neural networks, most notably convolutional neural networks (CNNs) for image tasks and recurrent neural networks (RNNs) for sequential data. The review also distinguishes radiomics, which extracts quantitative features from medical images, from the broader AI field.
Early-stage lung cancer is closely linked to lung nodules, which appear on CT scans as rounded or irregular opacities measuring up to 3 cm in diameter. While most are benign, some progress to malignancy. The volume of nodules discovered through increased screening creates a major workload challenge, and manual image reading is prone to missed diagnoses due to fatigue and inter-observer variability. AI tools address this at three stages: detection, segmentation, and classification.
Detection: Khosravan et al. proposed S4ND, a single 3D CNN with dense connections trained end-to-end for lung nodule detection, achieving a sensitivity of 95.2% in a single feedforward pass without additional post-processing. Kim et al. demonstrated that an AI-based tool improves the performance of both radiologists and pulmonologists when estimating malignancy risk for indeterminate pulmonary nodules on chest CT. Other detection models include 3D CMixNet (94% sensitivity, 91% specificity on LIDC-IDRI/LUNA16 datasets), U-net++ (94.2% sensitivity on 1,186 nodules), and SCPM-Net (89.2% average sensitivity on 888 CT scans).
Segmentation: After detection, precise delineation of nodules from surrounding pulmonary parenchyma is needed. Manual segmentation is time-consuming and highly variable between observers. Ronneberger et al. developed U-Net, a CNN architecture for biomedical segmentation that enables fine pixel-level outlining of nodule boundaries. Bhattacharyya et al. built on this with DB-NET, a weighted bidirectional feature network that improved U-Net performance specifically for ground-glass nodules, cavitary nodules, and small nodules. Li et al. achieved 99.02% accuracy with REMU-Net on 1,487 CT images.
Classification: The final step classifies nodules as benign or malignant, solid or sub-solid, and by specific subtype. Guo et al. proposed SAACNet, a 3D segmentation attention network integrating asymmetric convolution combined with a gradient boosting machine (GBM), evaluated on the LUNA16 dataset of 888 CT scans. It achieved 95.18% classification accuracy and an AUC of 0.977. Additional models include 3D SE-ResNet (AUC = 0.92 for classifying ground-glass nodules as invasive adenocarcinoma), DenseNet (92.4% accuracy), VGG-16/MobileNet/ResNet50 (92-95% classification accuracy), and YOLOv3/Faster-RCNN/SSD for detection (93-94% accuracy).
Accurate classification of lung cancer subtypes is clinically essential because treatment strategies differ dramatically between non-small cell lung cancer (NSCLC) and small cell lung cancer (SCLC), and between NSCLC subtypes such as lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). NSCLC accounts for approximately 80-85% of all lung cancers. Histopathologic confirmation remains the gold standard, but manual slide reading is time-consuming and susceptible to fatigue-driven errors.
Whole-slide image classification: Yu et al. used histopathology whole-slide images of LUAD and LUSC patients from The Cancer Genome Atlas (TCGA) to extract image features and employed regularized ML methods to differentiate short-term from long-term survivors. Their results demonstrated that automatically extracted image features can predict lung cancer prognosis. Coudray et al. trained Inception v3, a deep CNN, on 1,634 randomly selected histopathological whole-slide images from TCGA to classify them as LUAD, LUSC, or normal lung tissue. Their model achieved an average AUC of 0.97, consistent with pathologist-level analysis.
Cytopathological applications: AI has also been applied to fine-needle aspiration specimens, analyzing bronchial secretions, sputum, bronchoalveolar lavage fluid, and needle aspirates. Gonzalez et al. trained a DL algorithm based on CNN to differentiate SCLC from large-cell neuroendocrine carcinoma using cytological and biopsy specimens from 40 patients, accurately identifying the majority of both tumor types despite the small dataset.
Spatial analysis of the tumor microenvironment: Wang et al. developed ConvPath, a CNN-based software tool for LUAD digital pathology that performs nuclei segmentation, classifies tumor cells, stromal cells, and lymphocytes, and converts pathology images into spatial maps. The tool achieved 92.9% overall classification accuracy in training and 90.1% in independent testing. Beyond classification, ConvPath enables analysis of the spatial organization of cells and their roles in tumor progression and metastasis, and the authors validated a prognostic model for personalized treatment planning.
Lung adenocarcinoma is driven by a series of driver mutations, including mutations in epidermal growth factor receptor (EGFR), kirsten rat sarcoma viral oncogene homolog (KRAS), and anaplastic lymphoma kinase (ALK) fusions. These represent potential therapeutic targets, and identifying them non-invasively through imaging rather than tissue biopsy alone could significantly streamline clinical workflows and reduce patient burden.
Predicting mutations from pathology images: Coudray et al. downloaded gene mutation data from TCGA matched with patient samples and trained a CNN model to predict the 10 most commonly mutated genes in LUAD. Six of these genes (STK11, EGFR, FAT1, SETBP1, KRAS, and TP53) could be predicted from pathological images alone, with AUC values ranging from 0.733 to 0.856. This demonstrated that histopathological features encode detectable signatures of underlying genomic alterations.
CT-based prediction: Wang et al. developed the Fully Automated Artificial Intelligence System (FAIS), which uses non-invasive CT images to predict EGFR gene mutations with AUC values ranging from 0.748 to 0.813, and demonstrated statistically significant associations with progression-free survival (PFS) of patients receiving EGFR-targeted therapy (log-rank p < 0.05). The FAIS operates without requiring manual annotation, highlighting its scalability for clinical use.
ALK fusion prediction: Song et al. trained and validated a DL model using CT images and clinical pathological information to predict ALK fusion status in 937 NSCLC patients, achieving an AUC of 0.8046. In a subset of 91 patients undergoing ALK-TKI therapy, ALK-positive patients experienced significantly longer PFS (16.8 vs. 7.5 months, P = 0.010). This kind of predictive capability could help oncologists identify candidates for targeted therapy before invasive molecular testing.
Three-dimensional reconstruction technology has become an increasingly important component of AI-assisted lung cancer surgery. By reconstructing 3D images from CT, MRI, and PET-CT data, surgeons obtain clearer and more intuitive models of thoracic lesions and surrounding structures. This is especially critical because 20-30% of patients have variations in pulmonary vasculature, making preoperative identification of anatomical variations essential for safe surgery.
AI-assisted anatomical identification: In a retrospective cohort study, thoracic surgeons using AI-assisted CT achieved an 85% accuracy rate in identifying anatomical variations, with a median identification time of just 2 minutes. AI-driven reconstruction enables surgeons to rapidly and accurately recognize anatomical patterns, providing practical value for planning complex procedures like segmentectomy. CT-based pulmonary broncho-vascular 3D reconstruction is particularly advantageous for patients with deep nodules and vascular anatomical variations.
Semi-automatic vs. fully automated tools: Traditional 3D reconstruction relies on semi-automatic tools such as Mimics, OsiriX, and 3DSlicer, which require significant expertise and time. Li et al. constructed a fully automated 3D reconstruction system based on 3D CNN that reduced operation time by 24.5 minutes for lobectomy (P < 0.001) and 20 minutes for segmentectomy (P = 0.007). Compared to Mimics, the AI system cut model reconstruction time by 14.2 minutes (P < 0.001) and also outperformed Mimics in model quality scores (P < 0.001).
The authors note that current AI-based 3D reconstruction systems focus primarily on pulmonary vessels and bronchi. No reported studies have addressed automatic 3D reconstruction for tissue structure changes after neoadjuvant therapy, which they identify as a future development direction. Looking further ahead, AI-guided robotic surgery systems, AI-assisted lung cancer biopsy and treatment robots, and automated surgical systems may become viable clinical tools.
Preoperative invasiveness prediction: Determining the pathological invasiveness of lung cancer before surgery remains clinically difficult because it can only be definitively evaluated after examining the pathological specimen. Onozato et al. enrolled 873 patients who underwent lobectomy or segmentectomy, extracted radiomic features from preoperative PET and CT images, and compared seven ML models plus an ensemble model (ENS) combining PET and CT features. All models achieved an AUC of 0.880 or greater in the training set, while ENS showed the highest mean AUC of 0.880 and accuracy of 0.804 in the test set.
Risk stratification: Zhou et al. proposed an Ensemble Multi-View 3D Convolutional Neural Network (EMV-3D-CNN) for risk stratification of lung adenocarcinoma, outperforming senior physicians with an accuracy of 77.6%. Lv et al. developed a deep learning model that achieved performance comparable to intraoperative frozen section analysis in determining tumor invasiveness, potentially informing decisions about the extent of surgical resection.
Immunotherapy response prediction: Immune checkpoint inhibitors (ICIs) targeting PD-1 or PD-L1 have become the mainstay therapy for NSCLC patients without targeted treatment options, though only 20-40% of patients benefit from these therapies. The KEYNOTE-042 trial established that higher PD-L1 expression correlates with greater immunotherapy benefit. Monaco et al. constructed a tri-variate linear discriminant model extracting metabolic parameter features from PET/CT images, achieving 81% sensitivity and 82% specificity for predicting PD-L1 expression. Yang et al. combined radiomics with laboratory and clinical data to develop a DL model predicting response to anti-PD-1/PD-L1 drugs with an AUC of 0.80.
The authors emphasize that clinical indicators and radiomics features play complementary roles in PD-L1 status prediction. Integrating PET/CT, genomic data, and clinical information through multimodal analysis could identify additional predictive factors and further improve model accuracy. This represents a promising direction for development in the immunotherapy space.
Compared to traditional prognostic prediction based on clinical features alone, AI-assisted approaches offer greater accuracy and efficiency by extracting patterns from imaging and pathology data that human observers may miss. The review highlights two key approaches: radiomic feature extraction from CT scans and quantitative/spatial analysis of immunohistochemistry (IHC) images.
CT-based prognosis: Trebeschi et al. trained a neural network to identify morphological changes in retrospective chest CT scans of stage IV NSCLC patients during follow-up, linking learned radiomic features with overall survival (OS). The model showed significant performance in predicting 1-year OS, with an average AUC of 0.69. Performance peaked during the first 3-5 months of treatment, reaching an AUC of 0.75. The AUC for predicting sustained clinical benefit (defined as 6-month PFS) was 0.67.
IHC-based prognosis: Conventional analysis of single-plex chromogenic IHC is quantitative but lacks spatial analysis capability. AI algorithms can perform both quantitative and spatial analysis of immune checkpoint expression. Guo et al. utilized DL to analyze IHC pathology images and constructed a prognostic prediction model. U-Net segmented tumor cells and tumor-infiltrating lymphocytes, while ResNet extracted prognostic features from IHC images. The model achieved AUCs of 0.90 for predicting OS and 0.85 for predicting relapse-free survival (RFS).
The contrast between these two approaches is notable. The CT-based model provides moderate prognostic value (AUC 0.69-0.75) using non-invasive imaging that is routinely collected during follow-up. The IHC-based model achieves substantially higher performance (AUC 0.85-0.90) but requires tissue specimens. Together, these results suggest that combining non-invasive and invasive data sources through multimodal AI models could yield even stronger prognostic tools.
The authors identify five specific limitations that must be addressed before AI can be fully integrated into clinical lung cancer practice. These are not vague concerns but concrete, well-documented barriers that recur across the studies reviewed in this paper.
1. Data quality and standardization: AI models require substantial training data, but clinical datasets often suffer from missing information, biases, and noise. Equipment differences, imaging techniques, and treatment protocols vary across hospitals, undermining data consistency. The authors advocate for collaborative, standardized datasets and benchmarks to reduce variability and improve model robustness. 2. Interpretability: ML and DL models in lung cancer research frequently function as black boxes, making it difficult for clinicians to understand decision-making processes. Explainable AI (XAI) is a growing field that the radiology community has flagged as a priority for developing safe and intelligible AI technologies.
3. Generalization: Unlike natural images, medical images exhibit significant distribution differences across vendors and institutions. A model trained on data from one scanner or hospital may perform poorly when applied to images from another. This vendor-dependent performance degradation is a major concern for clinical deployment. 4. Ethics and privacy: Patient data used for predictions must be secured, regulated, and protected. AI should function as a diagnostic aid rather than a standalone tool, and potential biases in models could lead to unfair treatment outcomes. 5. Legal liability: The current legal ambiguity around autonomous AI in healthcare creates unresolved liability questions that could hinder adoption. Advocates need to push for legislative clarity that addresses liability without stifling innovation.
Future directions: The authors envision several developments. As algorithms improve and training data accumulate, AI will help detect lung imaging abnormalities with greater precision. AI-driven personalized treatment plans based on genetic information, clinical characteristics, and treatment response are on the horizon. Real-time treatment monitoring using sensor technology and AI could enable dynamic therapy adjustments. AI-assisted surgery, combining robotic systems with virtual reality, may reduce surgical risks and complications. Ultimately, the authors conclude that AI will not replace doctors but will complement their expertise, and the future lies in harnessing synergies between AI and healthcare professionals for patient-centered care.