Predicting pathological highly invasive lung cancer from preoperative [¹⁸F]FDG PET/CT with multiple machine learning...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Sublobar resection is effective for lung cancer but risks local recurrence in highly invasive cases.
• Accurate preoperative prediction of invasiveness is crucial for optimal treatment selection.
• Traditional methods like consolidation tumour ratio (CTR) have limitations in assessing invasiveness.

Purpose of the Study:

• To develop and validate machine learning models for predicting pathological invasiveness of lung cancer.
• To utilize preoperative [18F]fluorodeoxyglucose (FDG) positron emission tomography (PET) and computed tomography (CT) radiomic features for prediction.
• To compare the performance of machine learning models against the consolidation tumour ratio (CTR).

Main Methods:

• Radiomics features were extracted from preoperative PET/CT images of 873 lung cancer patients.
• Seven machine learning models and an ensemble (ENS) were evaluated using 100 iterations.
• Nested cross-validation was employed to assess calibration, clinical usefulness, and compare with CTR.

Main Results:

• Combined PET/CT radiomic features yielded an Area Under the Curve (AUC) of ≥0.880 in the training set.
• The ensemble model (ENS) achieved the highest mean AUC of 0.880 in the test set, with 80.4% accuracy.
• The model demonstrated high discriminative ability (AUC 0.882) and good calibration, outperforming CTR by over 8%.

Conclusions:

• Machine learning models integrating preoperative [18F]FDG PET/CT radiomics can reliably predict pathological invasiveness in lung cancer.
• The developed model offers high discriminative ability and stability for quantitative risk assessment.
• This approach shows significant potential to enhance preoperative lung cancer staging and treatment planning.