Abstract 7452: MAGNETICA: Machine-augmented generation of novel enhanced tomographic imaging for tumor insight and cancer assessment

Abstract Background: Quantitative features extracted from radiology images can reveal tumor characteristics beyond what the human eye can detect. While promising for identifying imaging biomarkers, the limited availability of patient 3D CT images in early-phase cancer trials is a major bottleneck. Recruiting patients with specific tumor types is difficult, leading to small, non-diverse datasets that lack statistical power for decisions on drug dosing, safety, and mechanism. Image augmentation could generate diverse, synthetic patient images to simulate tumor profiles, model disease progression, and increase biological variation while preserving patient privacy. However, current models do not meet these goals. To address this, we introduce MAGNETICA, a novel framework that integrates clinical features with anatomical structures for synthetic 3D CT image generation. Methods: We used the NSCLC-Radiomics (Lung1) dataset - which includes pretreatment CT scans, manual 3D gross tumor volume delineations by radiation oncologists, and clinical outcome data from 422 NSCLC patients, 36% were squamous histology and 68% were stage III - to train our model. A Variational Autoencoder (VAE) processed 3D CT volumes, preserving anatomical structures while compressing them into latent features, which were input into a Latent Diffusion Model (LDM) incorporating key clinical characteristics into its latent space. Clinical variables—including age, gender, tumor stage (T, N, M), histology, and survival time were transformed into textual descriptions using a sentence transformer model. These embeddings were integrated into the LDM to condition image generation. To ensure anatomical consistency, ControlNet was used to condition the generation on organ and lesion maps, producing realistic, patient-specific CT images with lesion labels. Synthetic images were validated with automated anatomical reconstruction scoring to ensure fidelity. Results: The model demonstrated strong alignment between synthetic and real 3D CT scans. The Mean Absolute Error between synthetic and real images was 0.12. For the radiomic feature “Total Energy, ” the Kolmogorov-Smirnov test had a statistic of 0.07, and the Kullback-Leibler divergence was 3.81×10-11, indicating no significant differences. Moreover, an automated anatomical annotation score validated that 98.9% of our VAE-generated CT synthetic images were high-quality, scoring 0.8 or higher. Conclusions: Consistent with our hypothesis, the synthetic 3D CT scans closely resembled real images, paving the way for capturing key disease characteristics. MAGNETICA advances synthetic imaging by bridging AI-driven generation and clinical needs. By augmenting early-stage trials with diverse, clinically relevant synthetic CT scans, this approach has potential to empower analyses that inform on dose selection, safety monitoring, and precision medicine. Citation Format: Hosein Barzekar, Mahyar Abbasian, Morteza Rezanejad, Mohammed Qutaish, Neal Sherkar, Deepti Nagarkar, Karen Stephansen, Scott Ogden, Lauren K. Brady, Brandon W. Higgs. MAGNETICA: Machine-augmented generation of novel enhanced tomographic imaging for tumor insight and cancer assessment [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 7452.