Radiological and biological dictionary of radiomics features: addressing understandable AI issues in personalized lung cancer; dictionary version LC1.0

Abstract

With late-stage lung cancer survival rates below 20%, there is a pressing need for AI tools that are both accurate and clinically interpretable. To address this, we developed the CT-Lung-RADS–based dictionary (LC1.0), which incorporates ExpertLRAD—a literature-based mapping of ten Lung-RADS descriptors to 68 radiomic measures, paired with SHAP analysis to translate model-important features into meaningful clinical terms. We assembled CT images and manual lesion masks for 977 patients from twelve TCIA cohorts (e.g., LIDC-IDRI, NSCLC-Radiomics, LungCT-Diagnosis); 499 had overall survival data (Class 1: Survival beyond 4 years (n = 264); Class 0: Death within 4 years (n = 235)), and the remaining 478 (LCTSC, Lung-Fused-CT-Pathology, NSCLC-Radiogenomics, QIN Lung CT, RIDER series, SPIE-AAPM Lung CT Challenge, TCGA-LUAD, and NSCLC-Radiomics-Genomics) were reserved for semi-supervised learning. All scans passed artifact QC, were reconstructed to 1 mm slices, resampled to 1 mm³ voxels, clipped to [–1000, 400] HU, min–max normalized, and 3D Gaussian-smoothed. Lesions were segmented into 3D Slicer by trained physicians and medical physicists and reviewed by a radiologist. From each lesion, we extracted 107 radiomic features, 2D/3D shape, first-order statistics, and higher-order textures, then applied min–max scaling. On the 499 labeled cases, we filtered features with five methods and trained five classifiers (AdaBoost, Bagging, Bernoulli/Complement Naïve Bayes, Decision Tree), optimizing hyperparameters via stratified five-fold cross-validation. Our top pipeline (Bernoulli Naïve Bayes + ANOVA F-test) achieved 0.88 ± 0.01 accuracy (AUC up to 0.89). ExpertLRAD accounted for roughly half of our ten most impactful SHAP-selected features—primarily within the shape, margin, and first-order intensity categories—while SHAP analysis identified novel proxies reflecting growth dynamics (Δ volume), zone-matrix variance (gray-level variance), and higherorder texture patterns (difference entropy, joint average, high–gray-level run emphasis). The framework links radiomics features to the Lung-RADS lexicon, confirming known findings, revealing new biomarkers, and providing a transparent, scalable bridge between AI outputs and clinical practice to foster trust and advance precision oncology.

Ähnliche Arbeiten