CT‐derived quantitative imaging metrics for pathological assessment of stage I pulmonary invasive adenocarcinoma: For the diagnostic examination, uAI Intelligent Imaging System was employed

Abstract For stage I lung infiltrative adenocarcinoma, pathological grade is key for treatment and prognosis. This research assessed how well a uAI‐based computer tomography (CT) imaging system differentiates pathological grades in stage I invasive lung adenocarcinoma. This study examined 260 patients with pathologically confirmed stage I pulmonary infiltrative adenomatous lesions, and the data were categorized into level 1 (32 cases), level 2 (182 cases), and level 3 (46 cases). Using a powerful uAI system for images, we extracted approximately 13 imaging parameters, including kurtosis, skewness, entropy, sphericity, major and minor diameters, fluid‐to‐part ratio, probability of integrity, and CT value statistics (maximum, minimum, average, median, and standard deviation). We then developed and evaluated different machine learning models, especially Logistic Regression, Support Vector Machine, XGBoost, and Random Forest were tested with a 5‐fold cross‐validation approach. Out of the 13 numerical parameters, CT_mean_HU showed the greatest variation (Kruskal‐Wallis test, ε 2 = 0.098, P < 0.001). Random Forest performed best, achieving 71.9% accuracy in 5‐fold cross‐validation and an average area under the curve (AUC) of 0.799. For different grades, the AUCs were 0.722 (Grade 1), 0.731 (Grade 2), and 0.848 (Grade 3). The most distinguishing features were CT_Mean_HU, CT_Median_HU, kurtosis, proportion of solid component, and entropy. In conclusion, quantitative CT parameters, particularly attenuation and texture features, show strong potential for non‐invasive grading of stage I lung invasive adenocarcinoma, and machine learning models based on these could aid clinical decisions.