Lung cancer is the leading cause of cancer death worldwide. One of the biggest challenges is that clinicians must interpret large, complex images (like CT scans) and patient data quickly and consistently. This review examines how artificial intelligence (AI) and machine learning (ML) can support earlier detection, more accurate diagnosis and staging, and clearer estimates of patient outcomes.
The paper evaluates recent methods for finding and characterizing lung nodules (small spots in the lungs that might be cancer), segmenting tumors (precisely outlining tumor boundaries), and predicting survival or treatment response. It also compares these AI tools with routine clinical practice to assess where they actually add value.
Importantly, the review also identifies the main barriers to real-world use, including data differences across hospitals, limited transparency of algorithms, and the need for external validation before these tools can be trusted in clinical settings.
TL;DR: A comprehensive review of how AI is being used to detect lung cancer earlier, diagnose it more accurately, and predict which treatments will work best.
Detection
AI for Lung Nodule Detection: Finding Cancer Before Symptoms Appear
Low-dose CT screening can catch lung cancer early, but radiologists must scan through hundreds of image slices per patient, looking for tiny nodules that might be just a few millimeters in size. It is easy to miss them, especially when they overlap with blood vessels or other structures.
AI systems trained on thousands of CT scans can now detect lung nodules with sensitivity rates exceeding 90% in many studies. Some systems can flag suspicious nodules that radiologists might miss, while also reducing false positives (spots that look concerning but turn out to be harmless).
The most promising approaches use deep learning, specifically convolutional neural networks (CNNs), which learn to recognize patterns in images much like the human visual system does, but they can process far more data far more consistently. Several AI tools have now received FDA clearance for clinical use as decision-support tools for radiologists.
TL;DR: AI systems can detect tiny lung nodules in CT scans with over 90% accuracy, catching cancers that human radiologists might miss.
Diagnosis
AI for Diagnosis and Staging: Not Just Finding, But Characterizing
Finding a nodule is only the first step. Doctors also need to determine whether it is benign or malignant, what type of lung cancer it is, and how far it has spread (staging). AI is being applied to all of these tasks.
Radiomics is a technique where AI extracts hundreds of quantitative features from medical images that are invisible to the human eye, such as texture patterns, shape irregularities, and intensity distributions. These features can help predict whether a nodule is cancerous without needing a biopsy.
For pathology, AI systems can analyze tissue slides (histopathology) to classify cancer subtypes, identify molecular markers, and even predict genetic mutations from the appearance of cells under a microscope. This is significant because treatment decisions often depend on these molecular characteristics, and traditional testing can take days or weeks.
TL;DR: AI can extract invisible features from images to predict if a nodule is cancerous, classify cancer subtypes, and even predict genetic mutations from tissue slides.
Prognosis
Predicting Outcomes: Who Will Respond to Treatment?
One of the most promising applications of AI in lung cancer is predicting patient outcomes. Machine learning models can integrate multiple data sources like imaging, lab results, genomic data, and clinical history to estimate survival, predict treatment response, and identify which patients are at highest risk of recurrence.
For example, AI models trained on CT scans taken before and during immunotherapy treatment can predict which patients will respond to the treatment, potentially sparing non-responders from ineffective therapy and its side effects. Other models can predict the risk of cancer coming back after surgery, helping guide decisions about additional treatment.
The review notes that while these predictive models show impressive accuracy in research settings, most have not yet been validated in diverse, real-world patient populations. This 'generalizability gap' is one of the biggest challenges facing AI in oncology.
TL;DR: AI can predict which patients will respond to treatment and who is at risk of recurrence, but most models still need real-world validation.
Challenges
Barriers to Real-World Use: Why AI Is Not in Every Clinic Yet
Despite promising results, the review highlights several major barriers preventing widespread clinical adoption of AI. First, most AI models are trained on data from a limited number of hospitals, and their performance often drops when applied to data from different institutions with different scanners, protocols, and patient demographics.
Second, many AI algorithms are 'black boxes.' They can make accurate predictions but cannot explain why they reached a particular conclusion. This lack of transparency makes clinicians hesitant to trust and act on AI recommendations, especially for life-or-death decisions.
Third, there is a lack of standardized evaluation frameworks. Different studies use different metrics, different datasets, and different comparison baselines, making it hard to know which AI tools are truly the best. The review calls for more prospective clinical trials, external validation studies, and regulatory frameworks to bridge the gap between AI research and clinical practice.
TL;DR: AI adoption is held back by limited generalizability across hospitals, 'black box' algorithms that cannot explain their reasoning, and lack of standardized evaluation.