Current clinical guidelines mandate routine evaluation of anaplastic lymphoma kinase (ALK) rearrangement in lung adenocarcinoma prior to ALK-targeted therapy initiation. This study aimed to develop and validate a non-invasive predictive model integrating deep learning radiomic (DLR) features from pre-treatment computed tomography (CT) images with clinical data to improve pretherapeutic ALK rearrangement prediction. We retrospectively analyzed 502 patients with histologically confirmed lung adenocarcinoma (153 ALK-positive, 349 ALK-negative), randomly split into training (80%) and validation (20%) cohorts. DLR features were extracted from pre-treatment CT images, and eight machine learning algorithms were compared. The optimal-performing algorithm was used to develop a combined clinical and deep learning radiomics (CDLR) model. Performance was evaluated via receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA). Gradient-weighted Class Activation Mapping (Grad-CAM) and Shapley Additive Explanations (SHAP) enhanced model visualization and interpretability. The support vector machine (SVM)-based DLR model yielded the best performance (training area under the curve (AUC): 0.971, 95% confidence interval (CI): 0.9528–0.9888; validation AUC: 0.877, 95% CI: 0.8071–0.9463). The CDLR model exhibited comparable efficacy (training AUC: 0.971, 95% CI: 0.9527–0.9889; validation AUC: 0.887, 95% CI: 0.8203–0.9530), with both significantly outperforming the clinical-only model (training AUC: 0.669, 95% CI: 0.6110–0.7273; validation AUC: 0.660, 95% CI: 0.5443–0.7757). Calibration analysis confirmed good agreement between predicted and observed outcomes. Our CT-based deep learning radiomics model holds promise for non-invasive detection of ALK rearrangements in lung adenocarcinoma, yet remains investigational and necessitates prospective multicenter validation before clinical implementation.
Li et al. (Thu,) studied this question.