Abstract 5775: Direct prediction of tumor cellular phenotypic composition from multi-parametric MRI in neuroblastoma murine models

Abstract

Listen

Translate article

Abstract Introduction. The clinical prognosis for children with high-risk neuroblastoma, a cancer of the sympathetic nervous system, remains poor. Non-invasive imaging biomarkers of treatment response are essential for precision medicine. Here, we evaluated a supervised machine learning model trained on co-registered MRI-histology maps to predict density maps of key neuroblastoma cell populations (undifferentiated, differentiating, apoptotic) directly from multi-parametric mpMRI in murine models of neuroblastoma. Methods. Tumour MRI parametric maps for spin-lattice relaxation time (T1, sensitive to undifferentiated cells), transverse relaxation rate (R2*, sensitive to red blood cells), and apparent diffusion coefficient (ADC, sensitive to cell density) were co-registered with digitized H&E slides (20x, 0.46μm resolution) from 11 Th-MYCN mice. Automated cell segmentation/classification into 6 categories allowed for the generation of density maps at MRI resolution. Classical ML models were trained to simultaneously predict 5 histological labels using a supervised regression approach on 9095 MRI voxels (n=4 mice), each represented by 5 normalized input features: R2*, T1, and ADC, along with median values of the pixels’ square neighbourhood (2px wide). Model performance was evaluated using 10-fold cross-validation and the negative mean-squared-error metric before application to remaining tumours. Results/Discussion. The random forest classifier achieved the lowest mean-squared-error and smallest variation during the 10-fold cross-validation and was selected as the best model. Extra trees, XGBoost and light-gradient-boosting machines also performed well. The similarity of the entire MRI-generated maps (for all 11 mice in the dataset) was compared to the histology maps using the mean squared error (MSE) and structural similarity index (SSIM) metrics. The model achieved an average MSE of 0.0037, 0.0089, 0.0016, 0.0014, and 0.0008 for cellular density, undifferentiated neuroblasts, apoptotic neuroblasts, differentiating neuroblasts and red blood cells density respectively. The SSIM indexes were 0.90, 0.88, 0.90, 0.88 and 0.92 for cellular density, undifferentiated neuroblasts’, apoptotic neuroblasts’, differentiating neuroblasts’ and red blood cells density respectively. Conclusion. This proof-of-concept study demonstrates that mpMRI can characterize neuroblastoma phenotypes and generate quantitative data previously accessible only through digital histology analysis, including cellular density maps. Validation in more diverse datasets could enable “non-invasive” or “virtual” biopsies, capturing entire tumour volume to enhance diagnostic precision, guide surgical planning, and refine response assessment. Citation Format: Konstantinos Zormpas-Petridis, Matthew Blackledge, Louis Chesler, Yinyin Yuan, Simon P. Robinson, Yann Jamin. Direct prediction of tumor cellular phenotypic composition from multi-parametric MRI in neuroblastoma murine models [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 5775.