Peritumoral and intratumoral radiomics for predicting visceral pleural invasion in lung adenocarcinoma based on preoperative computed tomography (CT)

Aim: To evaluate the prediction of peritumoral and intratumoral radiomics for visceral pleural invasion (VPI) in lung adenocarcinoma cancer (LAC) based on preoperative computed tomography (CT) radiomics.

Materials and methods: In total, 350 patients with LAC confirmed by surgery pathology were enrolled in XXX hospital, including 281 VPI negative patients and 69 VPI positive patients, were divided into the training cohort (n = 280) and validation cohort (n=70) at random with a ratio of 8:2. We extracted the radiomics features from the 3 region of interest (ROI), including gross tumor volume (GTV), the gross peritumoral tumor volume (GPTV) and the gross volume of the tumor rim (included the outer 4 mm of the tumor and 4mm of the tumor adjacent lung tissue on either side of the tumor contour boundary, GTR).The maximal redundancy minimal relevance (mMRM) algorithm and the least absolute shrinkage and selection operator (LASSO) was performed to reduce feature dimensionality and the radiomics score (Rad score) of the best radiomics model was combined with CT morphological characteristics with statistical significance in the univariable analysis to construct the combined model. The performance of the models was evaluated based on receiver operating characteristics (ROC) curve, calibration, and clinical usefulness. DeLong's test was used to assess differences in area under curve (AUC) between different models.

Results: There were no statistically significant differences in patient's gender, age, and BMI between the VPI positive group and VPI negative group (all p＞0.05). There were statistically significant differences in the tumor maximum diameter, tumor CT image type, vacuole sign, and pleural indentation sign between the VPI positive group and VPI negative group (all p < 0.05). The models of radiomics of GTV, GPTV, and GTR showed high predictive value in the training cohort (All AUC > 0.75). Compared with GTV, GTR radiomics models, the GPTV radiomics model constructed via the logistic regression (LR) method exhibited better prediction performance with the AUCs of 0.819, 0.827; accuracy of 0.757,0.743; sensitivity of 0.800,0.786; specificity of 0.747,0.732 in the training and validation cohorts, respectively. The LR model of GPTV radiomics was defined as the optimal model for predicting VPI, since its excellent performance in both ROC, calibration curve and decision curve analysis (DCA).

Conclusion: Preoperative CT-based radiomics models can predict VPI in patients with LAC; the LR algorithm combined the GPTV radiomics was the optimal choice, demonstrating high sensitivity, specificity, accuracy and clinical usefulness.