QATAB-SCLC: AN EXPLAINABLE QUALITY-AWARE MULTIMODAL PET/CT–CLINICAL MACHINE LEARNING FRAMEWORK FOR SMALL CELL LUNG CANCER PREDICTION

Abstract

Small Cell Lung Cancer (SCLC) is one of the most aggressive forms of lung cancer, characterized by rapid progression, early metastasis, and poor survival rates, making early and reliable prediction essential for improving clinical outcomes. Although recent advances in artificial intelligence have significantly enhanced lung cancer diagnosis, most existing studies primarily focus on classification or segmentation tasks using single-modality imaging and provide limited interpretability and prediction reliability. To address these limitations, this study proposes QATab-SCLC, an explainable Quality-Aware Multimodal PET/CT–Clinical Machine Learning Framework for SCLC probability prediction. The proposed framework integrates CT, PET, fused PET/CT images, and clinical information within a unified multimodal learning architecture. Initially, an image quality assessment module quantifies image reliability using statistical and texture-based quality descriptors. Handcrafted radiomic features and deep semantic representations extracted using a pretrained ResNet18 network are subsequently fused with clinical variables to construct comprehensive multimodal feature representations. After preprocessing and Mutual Information-based feature selection, two complementary tabular learning models, TabPFN and TabM, independently estimate SCLC probabilities. Their outputs are dynamically combined through a Quality-Aware Adaptive Fusion Network, which determines optimal fusion weights based on image quality and prediction uncertainty to generate robust and well-calibrated probability estimates. To enhance transparency, SHapley Additive exPlanations (SHAP) are incorporated to provide both global feature importance and patient-specific decision explanations. Experimental evaluation demonstrates that the proposed framework consistently outperforms existing machine learning and deep learning approaches across discrimination, calibration, and probability prediction metrics, including ROC-AUC, PR-AUC, Brier Score, Log Loss, RMSE, and Expected Calibration Error. The proposed QATab-SCLC framework offers a reliable, interpretable, and clinically applicable decision-support system for early SCLC probability prediction, supporting more informed diagnosis and personalized treatment planning.