Abstract Background: Non-small cell lung cancer (NSCLC), responsible for the majority of lung cancer-related deaths, displays prognostically meaningful intratumoral heterogeneity on CT imaging that is frequently overlooked by qualitative assessment. To advance quantitative risk stratification in precision oncology, we develop a deep learning framework to derive CT-based radiomic phenotypes in EGFR-mutated NSCLC and evaluate their association with overall survival. Method: A cohort of 130 NSCLC patients with expert-annotated tumor masks from the Stanford Radiogenomics Dataset was analyzed. To improve radiomic stability and reduce sampling variance, 1,300 augmented 2D tumor slices were generated using controlled augmentation protocols. For each patient, slices with maximal tumor burden were selectively curated to emphasize intratumoral heterogeneity in subsequent modeling. High-level morphological descriptors were extracted through a pre-trained ResNet50, and the resulting embeddings were projected into a lower-dimensional manifold via Parzen-Rosenblatt Isometric Mapping (PR-Isomap), preserving nonlinear geometric relationships linked to heterogeneous growth dynamics. Gap statistics determined the optimal latent dimensionality. These low-dimensional vector biomarkers were then used to classify survival outcomes. Results: PR-Isomap-derived embeddings demonstrated pronounced stratification of patient survival cohorts, underscoring the method’s capacity to encode clinically salient heterogeneity within tumor phenotypes. A ResNet50 backbone was optimized over 200+ training epochs using an Adaptive Moment Estimation-based optimizer with a learning rate initialized at 10-4 and adaptively decayed to ensure stable convergence. Model training employed a cross-entropy loss function to enhance discriminative capability across survival categories. The resulting classifier attained a 97% slice-level accuracy, exhibiting robust generalization across all augmentation regimes. Notably, model discrimination of survival endpoints was attributable to biologically meaningful tumor-intrinsic structural patterns, rather than extraneous anatomical noise, owing to segmentation-informed slice curation that systematically excluded non-tumoral confounders. Conclusion: Integrating deep learning-based CT-derived heterogeneity signatures with PR-Isomap embeddings enhances survival stratification in NSCLC. These imaging-driven, low-dimensional Deepomics features provide clinically actionable prognostic biomarkers, enabling earlier identification of high-risk patients and informing precision treatment planning. Citation Format: Joel Thomas, Blake Gilbert, Harmen Siezen, Lan Ma, Jimmy J Azarnoosh, Bardia Rodd. Leveraging PR-isometric guided deep learning to decode CT-based tumor heterogeneity and enhanced NSCLC prognosis abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 2792.
Thomas et al. (Fri,) studied this question.