Predicting benign, preinvasive, and invasive lung nodules on computed tomography scans using machine learning.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Lung nodules are frequently detected on computed tomography (CT) scans.
• Accurate differentiation of benign nodules from malignant or preinvasive adenocarcinoma subtypes is crucial for patient management.
• Current diagnostic methods can be invasive and carry risks.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) algorithms in classifying lung nodules.
• To determine if CT images alone are sufficient for differentiating benign, preinvasive, and invasive adenocarcinoma subtypes.
• To compare the performance of support vector machine (SVM) and convolutional neural network (CNN) models.

Main Methods:

• A dataset of chest CT scans with pathologically confirmed lung nodules was utilized.
• The dataset was randomly partitioned into training (70%), internal validation (15%), and independent test (15%) sets.
• Two ML models, SVM and CNN, were developed, trained, and validated for nodule classification.
• Receiver operating characteristic (ROC) analysis was employed to assess model performance.

Main Results:

• Both SVM and CNN models demonstrated varying degrees of success in classifying different lung nodule categories.
• The CNN model generally exhibited higher sensitivity compared to the SVM model, though with lower specificity and accuracy.
• Micro-average area under the curve (AUC) values reached 0.93 for SVM and 0.94 for CNN in differentiating benign/preinvasive from invasive adenocarcinoma.

Conclusions:

• ML algorithms, particularly CNNs, show potential for classifying lung nodules from CT images.
• These algorithms can reasonably differentiate between benign, preinvasive, and invasive adenocarcinoma subtypes.
• Further development could lead to improved, less-invasive diagnostic capabilities in lung cancer detection.