Information Fusion-Based Model for Lung Nodule Characterization

Abstract

Lung cancer continues to be one of the leading causes of cancer-related mortality worldwide, primarily due to the challenges of late diagnosis and the complexities involved in differentiating between benign and malignant pulmonary nodules. This thesis introduces a novel framework for lung nodule characterization that integrates information fusion techniques, combining deep learning representations with handcrafted radiomic features. The goal is to enhance classification performance, model interpretability, and clinical reliability. The proposed methodology employs a fusion strategy that blends deep features extracted from CNNs with shallow features, such as texture, shape, and intensity descriptors. These features are integrated at multiple levels of the network architecture to evaluate the effects of various fusion stages on performance. The system is rigorously assessed using the LIDC-IDRI dataset through extensive experimentation, including ablation studies, feature selection, and model explainability analyses utilizing Grad-CAM and SHAP. The results indicate that the integration of handcrafted and deep features leads to improved classification performance across different architectures, with consistent enhancements in AUC metrics. Additionally, the explainability analyses highlight that handcrafted features - especially shape descriptors - play a significant role in predicting malignancy. Overall, the findings suggest that this fusion-based approach not only boosts predictive accuracy but also enhances model transparency, providing a more reliable and interpretable tool for computer-aided lung cancer diagnosis. This research aligns with global efforts to promote early detection and ensure equitable access to healthcare, contributing to the development of AI-powered diagnostic tools that aid in informed clinical decision-making and advance the objectives of sustainable healthcare.