Lung cancer remains among the world’s most frequently diagnosed and deadly cancers. Although screening programs and targeted therapies have improved outcomes for some patients, major challenges persist. Early-stage disease can be difficult to distinguish from benign abnormalities, tumors can vary dramatically between patients, and the same tumor may evolve during treatment. Conventional clinical workflows often depend heavily on expert interpretation and incomplete snapshots of disease biology. A scan may reveal the shape of a lesion but not fully explain its molecular behavior; a biopsy may identify cancer cells but miss important differences between regions of the tumor; and a blood test may capture circulating signals that are invisible in tissue. Machine learning offers a way to combine these partial views and identify patterns that may be too subtle, multidimensional, or time-dependent for unaided human analysis.

The review was prepared by researchers from Shanghai Jiao Tong University, Qilu Hospital of Shandong University, Chongqing Medical University, and related institutions. The team organizes the field around a simple but powerful chain: data characteristics determine model selection, model performance determines the reliability of predictions, and reliable predictions must ultimately demonstrate clinical value. Five major data sources form the foundation of this framework. Medical imaging contributes information about tumor size, shape, density, texture, location, and changes over time. Multi-omics data—including genomics, transcriptomics, proteomics, and metabolomics—describe the molecular programs associated with tumor development and treatment response. Liquid biopsy can provide minimally invasive signals from circulating tumor DNA, RNA, proteins, or cells. Digital pathology captures cellular architecture at microscopic resolution, while clinical records provide demographic, physiological, treatment, and outcome information.

Each data type presents a different computational challenge. Imaging data are often high-dimensional and spatially structured, making convolutional neural networks and other deep-learning architectures useful for detecting features across pixels or three-dimensional scans. Digital pathology images can contain billions of pixels, requiring systems that divide slides into smaller regions before learning how local cellular patterns relate to a patient’s diagnosis or prognosis. Omics datasets, by contrast, may contain thousands of molecular variables but relatively few patient samples, creating a high risk of overfitting. Traditional machine-learning methods, feature selection, regularization, and dimensionality-reduction techniques can be valuable in such settings because they constrain the model and make its predictions more stable. Clinical data may include missing values, inconsistent terminology, and irregular time points, requiring specialized preprocessing and models capable of handling longitudinal information.

The article compares several generations of machine-learning approaches. Traditional methods such as logistic regression, support-vector machines, random forests, and gradient-boosting algorithms can perform well when datasets are moderate in size and features have been carefully defined. They are often easier to validate and interpret than more complex systems. Deep learning can learn representations directly from raw images, pathology slides, or other unstructured data, reducing the need for manual feature engineering. However, deep models generally require large, diverse, and consistently labeled datasets. Multimodal fusion methods attempt to combine information from different sources, either by integrating features early in the computational pipeline, merging model outputs at a later stage, or using architectures that learn relationships between modalities. These approaches can capture complementary signals, but they also face the problem of missing or poorly aligned data.

Foundation models represent another emerging direction. Trained on very large datasets, these models can learn general biological or visual representations and then be adapted to particular lung cancer tasks with less task-specific data. In principle, a foundation model trained on broad medical images or pathology material could be fine-tuned for nodule classification, tumor segmentation, subtype recognition, or treatment-response prediction. Yet the review emphasizes that scale alone does not guarantee clinical reliability. Training data may reflect one hospital, one scanner type, one population, or one style of clinical documentation. A model can therefore appear highly accurate in development while failing when transferred to a different institution. External validation, calibration, transparent reporting, and continuous monitoring are essential before such systems can influence patient care.

In early detection and diagnosis, multi-source machine learning could help clinicians distinguish malignant nodules from benign findings, prioritize patients for further testing, and identify cancers that might otherwise be overlooked. Imaging models can analyze subtle radiological patterns, including texture and spatial relationships that are difficult to describe using conventional measurements. When imaging is combined with clinical information, smoking history, laboratory results, or molecular signals from blood, the resulting prediction may be more informative than any single source alone. Similar strategies could support pathological diagnosis by linking tissue morphology with molecular subtypes and clinical outcomes. The aim is not simply to automate a radiologist’s or pathologist’s work, but to provide additional evidence, reduce variation, and help specialists focus attention on ambiguous or high-risk cases.

Treatment selection is another major area of opportunity. Lung cancer includes biologically distinct diseases that can respond very differently to surgery, chemotherapy, radiotherapy, targeted drugs, or immunotherapy. Machine-learning models can search for associations between molecular alterations, imaging features, pathological characteristics, treatment histories, and outcomes. These analyses may help estimate the probability that a patient will benefit from a particular therapy or develop resistance. Repeated measurements also make it possible to track disease dynamically. Changes in circulating tumor DNA, radiological appearance, or laboratory indicators could be analyzed over time to detect treatment response earlier than traditional assessments. Such systems could support adaptive treatment strategies, although the review stresses that predictions must be tested in prospective clinical studies rather than accepted solely on the basis of retrospective datasets.

Prognosis is similarly moving from a single end-of-treatment estimate toward continuous risk assessment. By integrating tumor biology, disease stage, treatment response, comorbidities, and follow-up information, machine-learning systems may identify patients at different risks of recurrence, progression, or treatment-related complications. This could allow surveillance schedules and supportive care to be tailored more precisely. However, the review identifies several barriers between promising algorithms and routine clinical use. Data standards remain inconsistent across hospitals; imaging protocols and pathology procedures vary; omics measurements can be expensive and technically heterogeneous; and patient records frequently contain missing or biased information. Multimodal models may also become less reliable when one data source is unavailable. In addition, clinicians and patients need to understand why a model produces a recommendation, especially when that recommendation affects an invasive procedure or life-changing therapy.

The researchers propose that future progress should focus on standardized data collection, adaptive fusion of complementary modalities, interpretable artificial intelligence, and prospective validation in real clinical environments. Interpretability does not necessarily mean reducing a complex model to a simple formula. It may involve showing which image regions influenced a prediction, identifying the molecular features associated with risk, quantifying uncertainty, or explaining how a patient’s current result differs from comparable cases. Fairness and privacy will also be central as hospitals connect large datasets and develop shared learning systems. Techniques such as federated learning could allow institutions to train models without transferring raw patient records, while robust governance frameworks could define how algorithms are audited, updated, and held accountable. The long-term vision presented in the review is a closed-loop lung cancer system in which early screening, diagnosis, personalized treatment, and prognosis are connected through continuously updated evidence. If that vision can be translated safely into practice, machine learning may help shift lung cancer management from experience-driven decisions toward precise, dynamic, and patient-specific care.

Subject of Research: Multi-source data-driven machine learning applications in lung cancer diagnosis, treatment, and prognosis.

Article Title: “Multi-source data-driven machine learning for lung cancer: diagnosis, treatment, and prognosis”

News Publication Date: 29 June 2026

Web References: https://doi.org/10.67704/ioe.2026.260006

References: Original review published in Intelligent Opto-Electronics, DOI: 10.67704/ioe.2026.260006.

Image Credits: Editorial Office of Opto-Electronic Journals Group.

Keywords: Lung cancer, machine learning, artificial intelligence, multimodal data, medical imaging, multi-omics, liquid biopsy, digital pathology, precision medicine, prognosis.