Pancreatic cancer preclinical research relies on multiple tumor models - different human cell lines grown as xenografts in mice - that have distinct biological characteristics. Being able to non-invasively distinguish between these models in vivo is valuable for understanding treatment response differences and validating which models best recapitulate specific aspects of human disease.
This study evaluated whether machine learning combined with Chemical Exchange Saturation Transfer (CEST) MRI could classify three distinct PDAC tumor models: Hs 766T, MIA PaCa-2, and SU.86.86 cell lines grown as xenografts in mice. CEST MRI is a specialized imaging technique that detects mobile protons in metabolites and proteins by exploiting magnetization transfer effects.
The study used a multi-modality imaging approach in 30 mice imaged twice, comparing CEST MRI to conventional T1-weighted MRI and Dynamic Contrast-Enhanced (DCE) MRI. The goal was to determine which imaging modality or combination best enables machine learning-based tumor type classification.
Chemical Exchange Saturation Transfer (CEST) MRI exploits the exchange of magnetization between water protons and exchangeable protons in metabolites, proteins, and glycosaminoglycans. By selectively saturating exchangeable proton pools, CEST MRI provides indirect information about metabolite concentrations and protein content without requiring exogenous contrast agents or radioactive tracers.
The CEST spectrum (also called the Z-spectrum) captures saturation effects at multiple frequency offsets from water, with specific peaks corresponding to different chemical species. The amide proton transfer (APT) signal around +3.5 ppm reflects mobile protein and peptide content, while signals at other offsets reflect glucose (glucoCEST) and other metabolites - all potentially altered in tumors with different metabolic profiles.
Acquiring full CEST spectra across multiple frequency offsets generates high-dimensional data per imaging voxel, making machine learning an appropriate analysis strategy: the algorithm can identify which combinations of spectral features most reliably distinguish between tumor types in ways that might not be obvious to human observers examining the spectra visually.
The machine learning pipeline combined Principal Component Analysis (PCA) for dimensionality reduction of the high-dimensional CEST spectral data with a k-Nearest Neighbors (kNN) classifier for tumor type prediction. PCA compresses the CEST spectrum into a smaller set of orthogonal components that capture the most variance, reducing noise while preserving discriminatory signal.
This PCA + kNN approach on CEST spectra achieved 87.5% classification accuracy across the three tumor models. This is a strong result for a three-class classification problem (chance level = 33%), demonstrating that CEST spectral patterns contain sufficient biological signal to distinguish between Hs 766T, MIA PaCa-2, and SU.86.86 xenograft models.
The use of kNN - a non-parametric classifier that classifies new observations based on similarity to training examples in feature space - is appropriate for the relatively small dataset (30 mice, imaged twice). More complex models would risk overfitting to the limited training data, whereas kNN's simplicity and interpretable distance-based reasoning is well-suited to this sample size.
Conventional T1-weighted MRI alone was insufficient for tumor type classification, as the signal intensities did not reliably distinguish between the three xenograft models. This finding confirms that standard anatomical MRI - which primarily reflects tissue water content and relaxation times - does not capture the functional and metabolic differences between PDAC cell lines that drive machine learning classification performance.
Dynamic Contrast-Enhanced (DCE) MRI, which tracks the kinetics of gadolinium contrast agent uptake and washout to characterize tissue vascularity and perfusion, also provided some classification ability. AUC values from DCE MRI-derived pharmacokinetic parameters could distinguish tumor types, reflecting real differences in the vascular biology of the three xenograft models.
The comparison across modalities demonstrates that functional and metabolic imaging (CEST, DCE) extracts complementary biological information beyond anatomical imaging. CEST's advantage over DCE is that it requires no exogenous contrast agent and provides metabolic rather than vascular information, potentially making it a safer and more informative first-choice modality for certain research questions.
The ability to non-invasively classify PDAC tumor models using CEST MRI + machine learning has practical value for preclinical drug testing. In longitudinal studies where multiple mice are imaged at multiple time points, being able to verify tumor identity and monitor subtype-specific characteristics without sacrificing animals enables more efficient and informative experimental designs.
Different PDAC xenograft models have different drug sensitivities, growth rates, and metabolic profiles - characteristics that are not always apparent from histology alone. The CEST-based classification approach could help researchers select the most appropriate preclinical model for testing a specific drug mechanism, improving the translational relevance of preclinical studies.
Extension to genetically engineered mouse models (GEMMs) of pancreatic cancer - which more accurately recapitulate the full tumor microenvironment including desmoplasia and immune components - would be the next logical validation step. If CEST MRI can distinguish GEMM subtypes, it could become a valuable longitudinal monitoring tool for studying disease progression and treatment response in the most sophisticated pancreatic cancer models available.