A Stacking Ensemble Learning Framework for Robust PET/CT Lung Cancer Multiclass Classification
Year: 2026
Type: conference-paper
Abstract: Lung cancer remains one of the leading causes of cancer-related mortality worldwide, making early and accurate diagnosis essential for improving patient outcomes. Positron emission tomography combined with computed tomography (PET/CT) provides complementary metabolic and anatomical information that can assist in lung cancer characterization. However, automated multiclass classification of lung cancer subtypes from PET/CT images remains challenging due to tumor heterogeneity and the high dimensionality of imaging features. This study proposes an extended hybrid machine learning framework for robust multiclass lung cancer classification using PET/CT images, integrating hybrid feature engineering, deep feature embeddings, and stacking ensemble learning. The proposed system integrates hybrid feature extraction, hybrid feature selection, and stacking ensemble learning to enhance classification performance. Radiomic descriptors and handcrafted metabolic features are extracted from PET/CT images to characterize tumor morphology, intensity distribution, and heterogeneity. In addition, deep feature embeddings are generated using a dual-stream multi-view convolutional neural network based on pretrained backbones to capture complementary PET and CT representations. A hybrid feature selection strategy combining mutual information, boosting-based importance, permutation importance, correlation filtering, and dimensionality reduction techniques is then applied to remove redundant features while preserving discriminative information. Finally, a stacking ensemble classifier consisting of Random Forest (RF), XGBoost, LightGBM, and TabNet is constructed, with logistic regression acting as the meta-learner. Experimental evaluation on five lung cancer classes demonstrates that the proposed framework achieves strong predictive performance. The stacking model obtains a test accuracy of 91.21% before feature selection, which further improves to 95.19% after feature selection, while reducing the feature dimensionality to 69 selected features. The model also achieves an area under the curve (AUC) of approximately 98.3%, indicating excellent class separability. These results demonstrate the effectiveness of integrating hybrid feature engineering with stacking ensemble learning for accurate and computationally efficient PET/CT lung cancer classification.
Institutions: Nahrain University, Obuda University
Language: English
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
OpenAlex ID: W7204157442