Artificial intelligence (AI) is advancing rapidly and holds great potential to transform lung cancer screening. By leveraging neural networks and other deep learning techniques, AI models can identify critical regions in scans and improve diagnostic accuracy. These tools can support earlier intervention, reduce late-stage diagnoses, and guide more personalized treatment.
Lung cancer remains the leading cause of cancer-related deaths worldwide, responsible for more deaths than breast, prostate, and colon cancer combined. Despite advances in treatment, one of the primary reasons for its high mortality rate is the late stage at which it is typically diagnosed. Approximately 75% of cases are detected in their advanced stages, with limited treatment options. Early-stage lung cancer often presents with little to no symptoms, making regular, proactive screening the most effective way to catch it in its early stages.
Low-dose computed tomography (LDCT) has emerged as the most effective tool for early detection. Studies such as the National Lung Screening Trial indicated a 20% reduction in lung cancer mortality through LDCT screening compared to traditional chest X-rays [1]. Despite its success, LDCT screening also has its limitations. Current screening guidelines primarily target individuals with significant smoking histories, however, there is growing evidence showing that lung cancer incidence is rising among populations with low or no tobacco exposure, particularly among women, Asian populations, and younger individuals. The current screening process not only limits access for at-risk individuals who fail to meet screening eligibility criteria but also fails to adapt to the complex, multifaceted nature of lung cancer risk. This trend underscores the need for more accurate and personalized screening strategies that reflect evolving risk factors and population trends using a data-driven approach.
Artificial intelligence (AI) has advanced rapidly in recent years and shows potential in addressing some of these gaps. By using available datasets and machine learning algorithms, AI systems can enhance the accuracy of lung cancer detection in LDCT scans, predict individual risk more precisely using clinical and demographic factors, and support radiologists in making clinical decisions. In the context of lung cancer screening, AI presents promising applications in detection of lung cancer, prediction of future lung cancer outcomes or occurrences, and predictive risk assessment.
Most AI models make use of machine learning (ML), a paradigm where computer algorithms learn patterns from large datasets in order to better interpret data and generate predictions. Machine learning can be broadly divided into two types: supervised learning and unsupervised learning.
In supervised learning, an algorithm is trained on a labeled dataset–data for which the correct output (such as disease presence, treatment response, or survival outcome) is already known–with the goal of being able to predict this output from new, unseen data. Internally, the algorithm contains a set of mathematical functions whose parameters are iteratively adjusted during training to minimize prediction error. This process allows the model to learn complex relationships between input features (e.g., imaging, clinical, or molecular data) and clinical outcomes. By contrast, unsupervised models do not require labels, and instead aim to learn patterns or groupings in data without explicit human guidance. Both supervised and unsupervised ML models show promising applications in lung cancer screening: supervised models can be used to identify malignancies or predict outcomes from screening results, and unsupervised models can help uncover new disease subtypes of risk patterns.
Many modern ML techniques rely on deep learning. Deep learning is a type of ML that uses artificial neural networks to understand complex patterns from data in a way that simulates the connections in the human brain. These neural networks are composed of multiple layers of interconnected neurons: each receives input signals, applies a mathematical operation to them, and produces an output signal. This multi-layered architecture allows deep learning models to capture and organize information hierarchically, enabling them to vastly outperform traditional models in a wide variety of tasks, including image-based recognition. As a result, deep learning models are especially well-suited for LDCT analysis.
As lung cancer screening increases, particularly among high-risk groups, AI can be used in improving detection of lung cancer. AI models have already been shown to perform as accurately [2] as radiologists in detecting nodules, and thus have the power to massively increase diagnostic accuracy and efficiency. One promising application is the ability of AI models to perform image segmentation, which is the task of separating an area of interest, such as a tumor, from the rest of an image [3]. As this technology improves, AI may also be able to better detect difficult-to-detect nodules or help filter out artifacts and increase the resolution of poor-quality scans, thus assisting radiologists in diagnostic throughput.
AI additionally shows promising applications in prediction of subtype, disease prognosis, or even future lung cancer occurrence from LCDT results. When trained on large datasets with years-long clinical followup data, deep learning models can learn complex predictive markers from LCDT images that predict long-term outcomes. For example, models have been developed to predict lung cancer occurrence 1-6 years in the future from LCDT radiographs, without the need for additional clinical data [4]. Models such as these demonstrate the power of AI to detect subtle, near incomprehensible patterns indicative of early malignant transformation before they become noticeable.
Finally, AI can play a critical role in risk assessment by integrating diverse data sources–such as age, race/ethnicity, geography, smoking history, imaging features, and genetic markers–to estimate the likelihood of developing lung cancer on an individual level [5]. AI can be used as a data analysis tool to uncover complex traits that can predispose individuals to be at higher risk, and use this knowledge to predict risk for individual patients. Incorporating such models in the clinic could identify high-risk patients years ahead of time, jumpstart the detection and treatment process, and save patients and providers valuable time and resources.
AI models, especially complex ones like deep learning algorithms, are often described as “black boxes” because their reasoning and internal workings are not easily interpretable by humans. A deep learning model, for example, may contain millions of parameters and contain complicated architecture, making it challenging for practitioners to interpret how an AI model arrived at its decision. Fortunately, there are a growing number of techniques being implemented aimed at making AI more interpretable.
One such method involves assessing model confidence, which reflects how certain the model is about its given prediction. This internal metric can be compared to a radiologist’s confidence in their diagnosis, offering a basis for comparison and validation. Another interpretability tool is the attention heatmap, which highlights regions of an image that the model focused on most when forming its prediction [6][7]. When overlaid with LDCT images, these heatmaps can be reviewed by clinicians to ensure the model is focusing on clinically relevant features, such as tumor morphology, borders, and surrounding tissue, thus improving trust in AI-assisted diagnoses.
While AI has shown substantial promise in lung cancer screening, several challenges must be addressed before it can be widely adopted in clinical practice. One key concern is algorithmic bias. Many AI models are trained on non-representative datasets collected at a limited number of academic medical centers, which can skew the model’s performance across demographic groups, including variations by race, sex, and smoking history. To address this, it is essential to train AI models on large, diverse, and demographically balanced datasets. This includes accounting for technical variability, such as differences in imaging equipment and radiographic technique. If it is infeasible or impossible to construct balanced datasets, techniques such as oversampling smaller classes or undersampling the majority class can be used during model development to best mitigate unbalanced data.
Beyond algorithmic bias, unbalanced datasets can hide poor model performance with misleading performance metrics. For example, for a dataset where 10% of samples are malignant and 90% of samples are not malignant, a model can achieve a test accuracy of 90% if it always predicts that a sample is non-malignant. In other words, on an unbalanced dataset, a model can achieve a high accuracy by simply “cheating,” rather than learning features that are truly predictive of malignancy. This underscores the importance of optimizing for and reporting multiple metrics, such as precision, recall (also known as sensitivity), F1 score, and ROC-AUC, which together offer a complete measure of a model’s performance and predictive capability. Thoroughly testing emerging AI models and reporting a variety of metrics ensures that they are well-suited to their clinical objective.
Another limitation involves the detection of small or poorly defined pulmonary nodules. Nodules located near the lung periphery or blood vessels have proven challenging to recognize accurately, even with advanced convolutional neural networks. These subtle features can be easily missed or misclassified [8], leading to false negatives and unnecessary follow-ups and biopsies due to false positives. Ways to address this may include curating high-quality datasets with rigorous annotation, as well as using state-of-the-art model architectures that can better differentiate nodule borders from surrounding tissue.
Overcoming these challenges is crucial in fully utilizing the clinical potential of AI, particularly in guiding more personalized and efficient screening strategies for diverse patient populations.
AI models are highly versatile and can assist with a multitude of objectives in lung cancer screening. At the imaging level, they can analyze raw LDCT images to assist with rapid detection and diagnosis. Additionally, they can be tailored to integrate relevant clinical metadata—such as age, sex, race and ethnicity, smoking history, and even genetics—to generate individualized risk predictions. They could even be modified to interface with other AI systems embedded within electronic health records (EHRs), enabling automated extraction of patient information and reducing burden for providers.
Moreover, AI models can support targeted screening strategies, taking into account race, ethnicity, habits, and even health data from wearables to alert patients who may be at higher risk to seek screening. Further, AI in the form of large language models such as Chat-GPT may play an important role in patient education, helping patients understand why lung cancer screening is important and alleviating concerns they may have about the process.
When combined with clinical, imaging, and behavioral data, AI models have the potential to not only provide personalized diagnoses and predictions, but to also uncover population-level, multi-modal prognostic subgroups that may inform new risk models and standards of care. As we work to expand access and participation in lung cancer screening, AI will be instrumental in enhancing throughput, improving diagnostic accuracy, and delivering timely, individualized information to both patients and providers.
References
[1] Reduced Lung-Cancer Mortality with Low-Dose Computed Tomographic Screening. (2011). New England Journal of Medicine, 365(5), 395–409. https://doi.org/10.1056/nejmoa1102873
[2] Ramli, P. N. M., Aizuddin, A. N., Ahmad, N., Hamid, Z. A., & Ismail, K. I. (2025). A Systematic Review: The role of Artificial Intelligence in Lung Cancer Screening in Detecting Lung Nodules on Chest X-Rays. Diagnostics, 15(3), 246. https://doi.org/10.3390/diagnostics15030246
[3] Primakov, Sergey P., et al. “Automated Detection and Segmentation of Non-Small Cell Lung Cancer Computed Tomography Images.” Nature Communications, vol. 13, no. 1, June 2022, p. 3423. www.nature.com, https://doi.org/10.1038/s41467-022-30841-3.
[4] AI for Early Detection of Lung Cancer | Mass General Brigham. (2024, September 5). https://www.massgeneralbrigham.org/en/about/newsroom/articles/ai-for-early-detection-of-lung-cancer
[5] S.k.b, Sangeetha, et al. “An Enhanced Multimodal Fusion Deep Learning Neural Network for Lung Cancer Classification.” Systems and Soft Computing, vol. 6, Dec. 2024, p. 200068. ScienceDirect, https://doi.org/10.1016/j.sasc.2023.200068
[6] Hammad, M., ElAffendi, M., El-Latif, A.A.A. et al. Explainable AI for lung cancer detection via a custom CNN on CT images. Sci Rep 15, 12707 (2025). https://doi.org/10.1038/s41598-025-97645-5
[7] Klangbunrueang, Rapeepat, et al. “AI-Powered Lung Cancer Detection: Assessing VGG16 and CNN Architectures for CT Scan Image Classification.” Informatics, vol. 12, no. 1, Feb. 2025, p. 18. https://doi.org/10.3390/informatics12010018.
[8] Cellina, M., Cacioppa, L. M., Cè, M., Chiarpenello, V., Costa, M., Vincenzo, Z., Pais, D., Bausano, M. V., Rossini, N., Bruno, A., & Floridi, C. (2023). Artificial intelligence in lung cancer screening: The future is now. Cancers, 15(17), 4344. https://doi.org/10.3390/cancers15174344
Jeffrey Velotta, MD, FACS, is a thoracic surgeon, clinical professor in the Department of Clinical Science at the Kaiser Permanente Bernard J. Tyson School of Medicine, and clinical assistant professor in the Department of Surgery at the University of California, San Francisco (UCSF) School of Medicine. Dr. Velotta’s research interests include improving lung cancer screening rates among all at-risk populations, as well as developing novel tools for earlier lung cancer detection.
Darren Chan is a researcher at Stanford University pursuing a Bachelor of Science in Biomedical Computation. His academic interests lie at the intersection of computer science, biology, and medicine, with a focus on developing computational tools to advance healthcare and biomedical research.
Barkha Trivedi is a researcher at the University of California, Los Angeles, pursuing a Bachelor of Science in Biology. With a strong interest in medicine and healthcare advocacy, she is passionate about addressing healthcare disparities and improving access to medical resources for underserved communities.
