Late-stage diagnosis dominates. Non-small cell lung cancer (NSCLC) accounts for about 85% of all lung cancers and is responsible for approximately 1.8 million deaths annually worldwide. A major driver of these poor outcomes is that 49-53% of patients are diagnosed at Stage IV - advanced or metastatic disease - where the five-year survival rate is often below 5%.
Early detection transforms outcomes. In stark contrast, patients diagnosed with localized, early-stage NSCLC can achieve five-year survival rates of 70-90% after surgical resection. This dramatic survival difference makes early and accurate diagnosis the most impactful intervention available.
Current diagnostic tools have real limitations. Tissue biopsy through bronchoscopy or needle aspiration remains the diagnostic gold standard but is invasive, carries complication risks, and is not feasible for all patients. CT screening detects nodules effectively but suffers from poor specificity - many benign conditions such as pneumonia, granulomas, and inflammatory masses can look identical to early cancer on imaging, leading to high false-positive rates and unnecessary procedures.
Blood-based biomarkers as a complementary tool. Serum tumor markers - proteins released by tumors or by the body's response to cancer - offer a minimally invasive, affordable, and repeatable alternative. Individually, these markers lack the accuracy needed for standalone diagnosis, but combined into a panel and analyzed with AI, they may provide a more reliable picture of what is happening in the lung.
A large, diverse cohort. The retrospective study enrolled 2,283 participants divided into four diagnostic groups: 1,339 with confirmed NSCLC, 313 with pneumonia, 260 with biopsy-confirmed benign lesions, and 371 with other benign lung masses. This design was intended to allow the model to learn to distinguish NSCLC from a realistic range of conditions that can mimic cancer on imaging.
Six serum biomarkers measured. Blood levels of six tumor markers were quantified using ELISA (enzyme-linked immunosorbent assay): Squamous Cell Carcinoma Antigen (SCCA), Carcinoembryonic Antigen (CEA), Cancer Antigen 125 (CA-125), Cytokeratin Fragment 21-1 (CYFRA21-1), Neuron-Specific Enolase (NSE), and Pro-Gastrin-Releasing Peptide (ProGRP). Each was chosen based on documented relevance to NSCLC diagnosis and differential diagnosis.
Why these specific markers. CYFRA21-1 shows high specificity for NSCLC. CEA and CA-125, while non-specific, improve accuracy when combined with other markers. SCCA is particularly associated with squamous cell carcinoma (a major NSCLC subtype). NSE and ProGRP are primarily associated with small cell lung cancer (SCLC), but their inclusion helps distinguish NSCLC from SCLC - a clinically critical differentiation.
Training and test split. The dataset was divided 70/30 into training and independent testing sets, with stratification to ensure proportional representation of all four diagnostic categories. The model was evaluated on the test set using accuracy, sensitivity, specificity, F1-score, and area under the ROC curve (AUC).
A custom deep neural network for six inputs. LungPanelNet is a feed-forward deep neural network that takes the six biomarker concentration values as inputs. Despite the small number of input features - just six numerical values - a deep learning architecture was chosen specifically for its ability to capture complex, non-linear interactions between biomarkers that simpler models cannot represent.
Two hidden layers with regularization. The network consists of an input layer, a first hidden layer with 128 neurons, and a second with 64 neurons. Each layer uses ReLU activation functions to enable non-linear decision boundaries, batch normalization to stabilize training, and a 50% dropout rate to prevent overfitting by randomly disabling half the neurons during each training pass.
Four-class output. The output layer produces predictions for four diagnostic categories: NSCLC, pneumonia, biopsy-confirmed benign lesions, and benign lung mass. This multi-class design is more clinically relevant than a simple cancer/no-cancer binary - what matters in practice is not just whether cancer is present, but what the condition actually is.
Benchmarked against traditional classifiers. LungPanelNet was directly compared against four conventional machine learning approaches: Naive Bayes, k-Nearest Neighbors (k-NN), Support Vector Machine (SVM), and Logistic Regression. All methods used the same dataset, ensuring a fair comparison of architectural complexity and feature capture capability.
High sensitivity for cancer, poor specificity for benign conditions. The model correctly identified 266 out of 268 cancer cases - a 99.3% sensitivity for the NSCLC class. However, it achieved this at the cost of misclassifying 170 non-cancer cases as malignant. The model failed entirely to correctly identify any case of biopsy-confirmed benign lesions or benign lung mass, with zero true positives for both of these classes.
Highly variable AUC across classes. ROC curve analysis confirmed the imbalance in performance. The best-performing class achieved an AUC of 0.80, others scored 0.71 and 0.66, and one class performed at AUC 0.51 - statistically indistinguishable from random guessing. The overall model AUROC in benchmarking was 0.676.
Better than conventional methods, but not by a wide margin. LungPanelNet's AUROC of 0.676 outperformed SVM (0.613), k-NN (0.605), and Naive Bayes (0.598), confirming that deep learning captures biomarker interactions better than traditional classifiers. It also outperformed logistic regression (0.641), suggesting non-linear feature relationships matter in this data.
Overfitting observed during training. Training and validation loss curves diverged after approximately epoch 8 - training loss continued to fall while validation loss plateaued and fluctuated. This classic overfitting pattern indicates the model began memorizing training data rather than learning generalizable patterns. Early stopping at epoch 8 would have produced the best generalizing version of the model.
Why the model is biased. The root cause of LungPanelNet's classification failure for benign conditions is clear: the dataset was dominated by cancer cases (1,339 out of 2,283 participants, or 59%). The model learned that predicting 'cancer' for most inputs would minimize its training error, since cancer was the most common outcome. This is a well-known pitfall in machine learning called majority class bias.
Planned solutions for future work. The authors identify two primary technical remedies. First, applying Synthetic Minority Oversampling Technique (SMOTE) to generate realistic synthetic data for the underrepresented pneumonia and benign classes, giving the model more examples to learn from. Second, implementing cost-sensitive learning - assigning higher penalties to misclassifying minority classes during training to force the model to attend to them.
Multi-center data as the longer-term fix. The study's single-center, retrospective design contributed to the imbalanced class distribution. Large, prospectively designed, multi-center studies inherently reduce selection bias and produce more balanced, representative datasets. Future validation will require this type of robust cohort to determine whether the approach can generalize across clinical settings.
Biomarkers alone may not be enough. The authors acknowledge that six serum markers, while informative, may be insufficient on their own to reliably distinguish clinically overlapping conditions. A multi-modal approach - combining biomarker data with CT-based radiomic features, smoking history, and clinical context - is likely needed to achieve the diagnostic accuracy required for real-world clinical use.
Not a screening tool for healthy people. The authors are explicit that LungPanelNet is not intended for primary screening of asymptomatic individuals. Its intended role is as a decision-support tool for patients who already have a radiological lung abnormality - such as an indeterminate pulmonary nodule found on CT - and need help determining whether it is likely malignant or benign.
A triage pathway. In the envisioned workflow, when a pulmonary nodule is identified on imaging, the model provides a probability score for malignancy. A high cancer probability score would trigger urgent referral for PET-CT or biopsy. A high benign probability score would support a watchful waiting approach with interval imaging, potentially sparing the patient from unnecessary invasive procedures.
Integration with clinical judgment is essential. The model's output is intended to serve as one input into a multidisciplinary decision-making process - not as a standalone diagnostic conclusion. The clinician's assessment, radiological features, patient history, and other laboratory data must all factor into the final management decision.
Missing data is a practical challenge. The model was validated assuming all six biomarkers are available. In routine clinical practice, some values may be missing or unavailable. The performance of the model with incomplete biomarker panels was not evaluated, representing an important gap that must be addressed before clinical deployment through imputation methods or minimum data requirements.
What this study achieved. LungPanelNet outperformed conventional machine learning methods on all metrics, establishing that deep neural networks are better suited than traditional classifiers for modeling complex interactions within a serum biomarker panel. The study provides a working architecture and a comprehensive analysis of both its capabilities and its current limitations.
What remains to be solved. The model's clinical usefulness is currently limited by its failure to reliably classify benign conditions - the exact capability needed to prevent false-positive diagnoses and unnecessary invasive workups. Addressing class imbalance through oversampling, cost-sensitive learning, and multi-center data collection is the essential next step.
Potential long-term impact. A validated, balanced version of LungPanelNet could fill a genuine gap in the lung cancer diagnostic pathway: providing a quick, inexpensive, non-invasive risk stratification tool that helps guide which patients with CT-detected nodules need urgent invasive evaluation and which can be safely monitored. This could reduce both missed cancers and unnecessary procedures across a large population of patients with indeterminate lung findings.