Bayesian Probabilistic Knowledge from Diameter Prior for Decision Fusion to Detect Lung Nodule Heterogeneity

Recent advances in deep learning (DL) have shown promising results in the identification of malignant lung nodules from computed tomography (CT) scans. However, conventional DL models primarily rely on spatial features and often lack clinical interpretability and the ability to incorporate domain-specific priors. To address these limitations, we integrate Bayesian diameter posterior probabilistic knowledge, allowing the model to leverage the well-established correlation between nodule size and malignancy likelihood. To enhance interpretability, we introduce a new deep-texture-shape (DTS) scoring scheme that offers oncologists a transparent, component-wise justification for malignancy prediction. Furthermore, relying solely on spatial features may overlook critical textural patterns, which play a significant role in malignancy assessment. To overcome this, we utilize local binary patterns (LBP), histogram of oriented gradients (HOG), and gray level co-occurrence matrix (GLCM) to extract rich textural features, capturing the fine-grained patterns that contribute to malignancy characterization. This multi-faceted approach not only improves prediction accuracy but also enhances the model’s robustness. The proposed method, probabilistic knowledge-based decision fusion (ProKDF), is evaluated on two public datasets: the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) and the International Society for Optics and Photonics and American Association of Physicists in Medicine (SPIE-AAPM). It achieves an F1-score of 87.48% (95% CI: 85.93%–89.03%) and an area under the receiver operating characteristic curve (AUC) of 92.56% (95% CI: 91.70%–93.42%) on LIDC-IDRI. Our findings suggest that integrating shape-based probabilistic knowledge and texture information enhances model performance and interpretability in detecting nodule heterogeneity.