Table 13.
Interpretation of proposed framework findings with respect to pathophysiology, radiomics and existing literature.
| Component/finding | Observations (proposed work) | Pathophysiological interpretation | Radiological significance | Alignment with related work |
|---|---|---|---|---|
| Entropy (Radiomics) | High positive correlation with malignancy | Reflects tumour heterogeneity, necrosis and disorganized cellular structure | Appears as irregular intensity variation in CT scans | Consistent with Zhang et al. (2023) and Adiraju et al. (2025) |
| GLCM contrast | Strong positive correlation | Indicates variation in tissue density due to aggressive tumour growth | Visible as heterogeneous texture patterns | Supported by radiomics-based studies on tumour heterogeneity |
| Surface-area-to-volume ratio | High values linked to malignancy | Represents spiculated and invasive tumour boundaries | Irregular, non-smooth tumour margins in CT | Aligns with clinical radiology findings of malignant nodules |
| Sphericity | Negative correlation with malignancy | Benign tumours tend to grow uniformly | Smooth, well-defined nodules | Consistent with standard radiological diagnostic criteria |
| Homogeneity | Negative correlation | Indicates uniform cellular structure in benign lesions | Even intensity distribution in CT scans | Supported by radiomics and imaging studies |
| HybridNET performance | 91.3% accuracy, AUC 0.93 | Captures both volumetric and contextual tumour characteristics | Improves detection of subtle malignancy cues | Outperforms 3D CNN (Alakwaa et al., 2017; Wu et al., 2023) |
| 3D + 2D fusion | Improved generalization | Combines spatial continuity with local texture | Enhances lesion characterization across slices | Similar trend reported in hybrid models (Li et al., 2024) |
| Segmentation (DSC 0.926) | Accurate boundary delineation | Captures tumour invasiveness and irregular growth | Important for staging and treatment planning | Comparable to advanced segmentation models (V-Net, U-Net variants) |
| Grad-CAM (SRO 83.5%) | High overlap with expert annotations | Model focuses on biologically relevant tumour regions | Visual validation of diagnostic regions | Aligns with XAI studies (Hammad et al., 2025) |
| SHAP feature attribution | Entropy, contrast most influential | Confirms biological relevance of radiomic features | Provides quantitative explanation of diagnosis | Supported by XAI frameworks (Fontes et al., 2024) |
| Dual XAI (Grad-CAM + SHAP) | Spatial + feature-level interpretability | Mimics clinical reasoning (where + why) | Enhances trust in AI decisions | Extends prior single-XAI approaches |
| Ablation study | Performance drops without 3D/radiomics | Each component captures distinct tumour properties | Confirms necessity of multi-modal analysis | Supports hybrid framework design principles |