Introduction

Pancreatic cancer is one of the most aggressive types of cancers. It is difficult to detect due to its asymptomatic presentation at early stages. Therefore, there is an unmet need for non-invasive imaging markers that help identify the aggressive sub-type(s) in a pancreatic lesion at an early time point or evaluate the efficacy of therapy in pancreatic cancer. In last few years, there are two major developments that can have significant impact in developing imaging biomarkers for pancreatic cancer: I) hyperpolarized metabolic Magnetic Resonance (MR) and II) application of Artificial Intelligence (AI) on imaging pancreatic cancer. One of the most commonly used imaging biomarkers are the conversion of hyperpolarized pyruvate to lactate and alanine.¹ Afterwards, the addition of Deep Learning (DL) algorithm is applied for AI to learn features from both metabolic and anatomical MR imaging modalities which has the potential for the early detection of pancreatic cancer, rather than simply monitoring tumor growth.²

Methods

Hyperpolarized 1-¹³C Pyruvate MRS was employed to study the metabolic processes in genetically engineered mouse (GEM) models (P48:Cre; LSL-KRAS^G12D (KC)) with pre-invasive pancreatic intraepithelial neoplasia (PanIN) precursor lesions and control animals (P48:Cre or WT C57BL/6) without pancreatic lesions.. The dissolution DNP (HyperSense, Oxford Instruments) operating at 3T was utilized to hyperpolarize 1-¹³C pyruvate. The ¹³C magnetic resonance spectra of hyperpolarized 1-¹³C pyruvate were acquired at 7T Bruker MRI scanner. These mice were imaged at different time points in their lifespan, at 14, 21 and 28 weeks. The biochemical changes of alanine transaminase (ALT) and lactate dehydrogenase (LDH) enzyme activity were assessed. Afterwards, advanced Deep Learning (DL) techniques are implemented to develop DL model to reveal hybrid biomarkers from MRI and metabolic imaging to predict early detection of pancreatic cancer. This model is developed based on advanced Bayesian deep learning techniques and multi-modal data integration to enable uncertainty measurements and learn features from both imaging modalities to consider to improve prediction accuracy. After training the model, the learned features from multiple modalities to identify any correlation between MRI and metabolic imaging are explored that may lead to the discovery of new hybrid biomarkers with predictive values for the early detection.

Results/Discussion

The alanine-to-lactate signal intensity ratio was found to decrease as the disease progressed from low-grade PanINs to high-grade PanINs (Figure 1). These results demonstrate that there are significant alterations of ALT and LDH activities during the transformation from early to advanced PanINs lesions. Furthermore, we demonstrated that real-time conversion kinetic rate constants (k_PA and k_PL) can be used as metabolic imaging biomarkers of pancreatic premalignant lesions (Figure 2). The appropriate DL combination features from MRI and metabolic imaging as complementary modalities that can lead to proper prediction of early detection in this KC GEM model.

Conclusion

Findings from this emerging DL and HP-MRS techniques can be potentially translated to the clinic for detection of pancreatic premalignant lesion in high-risk populations.

Acknowledgements

This research was funded in part by a grant from Pancreatic Cancer Action Network (PANCAN; 16-65-BHAT) (PKB, FM); Cancer Prevention and Research Institute of Texas (CPRIT; RP180164) (PKB) by Institutional Research Grants (PKB) and a Startup grant (PKB) from MD Anderson Cancer Center; by grants from the US National Cancer Institute (U01 CA214263, U54 CA151668 and R21 CA185536, R01 CA218004; and 1P50 CA221707-01). This work also was supported by the National Institutes of Health/NCI Cancer Center Support Grant under award number P30 CA016672.