Cross-Population Generalizability of Lung Cancer Radiomics and Dosiomics: Implications for AI Deployment in a South Asian Cohort

BACKGROUND: The integration of Artificial Intelligence (AI) into radiation oncology offers 
transformative potential for personalized medicine, particularly in predicting radiation-induced lung 
injury. However, the generalizability of AI models trained on standardized Western datasets to Low
and-Middle-Income Countries (LMICs) remains unproven. Differences in patient demographics, 
disease pathology, and imaging protocols may create a significant "domain shift," rendering imported 
models unreliable. This study presents the first comprehensive forensic audit of radiomic and dosiomic 
domain shift between a public Western reference cohort and a multi-center South Asian clinical cohort. 
METHODS: A total of 597 patients with non-small cell lung cancer (NSCLC) were analyzed, 
comprising a reference cohort of 422 patients from the NSCLC-Radiomics public dataset (The Cancer 
Imaging Archive) and a local test cohort of 175 patients retrospectively collected from two tertiary 
centers in Bangladesh: Square Hospital (n=121) and Ahsania Mission Cancer and General Hospital 
(n=54). To address the challenge of heterogeneous legacy data, a custom forensic DICOM processing 
pipeline was developed, including a universal folder organization algorithm and a geometric force
match strategy that successfully recovered 92.5% of local cases with coordinate system 
incompatibilities. Feature extraction was performed to capture both anatomical texture (radiomics) and 
dose distribution texture (dosiomics). Statistical analysis employed Principal Component Analysis 
(PCA) to visualize population shifts and Explainable AI (SHAP) to identify specific features driving 
divergence. 
RESULTS: The dosiometric audit revealed high stability in treatment quality across centers. Despite 
using different treatment planning systems (Varian Eclipse vs. Elekta Monaco), both local hospitals 
achieved consistent target coverage, with mean D95 values of 60.5 ± 2.1 Gy (Square) and 59.8 ± 3.4 
Gy (Ahsania) (p > 0.05), comparable to international standards. In stark contrast, the radiomic analysis 
exposed a profound domain shift. A Random Forest classifier trained to distinguish between Western 
and South Asian domains achieved an Area Under the Curve (AUC) of 1.00 on the test set, confirming 
that image features contained a strong institution-specific signature. Statistical analysis identified 
significant divergence in texture features, with 'Texture: LowGrayLevelEmphasis' showing a fold 
change of 0.23 (p < 0.001) between cohorts. SHAP analysis further revealed that geometric shape 
descriptors (e.g., Sphericity) and scanner-specific noise profiles were the primary drivers of this 
divergence. 
CONCLUSION: This thesis provides empirical evidence that direct importation of Western AI models 
into the Bangladeshi clinical workflow carries a significant risk of failure due to severe radiomic 
incompatibility. While the physics of treatment delivery is robust and transferable, imaging data 
requires local calibration. The identified domain shift necessitates a transition from passive AI adoption 
to active data curation and the development of sovereign models grounded in local reality.