# Radiomics in sporadic microsatellite instable, mismatch repair deficient and Lynch syndrome-associated pancreatic ductal adenocarcinoma: a pilot study

## TL;DR

This pilot study explores whether radiomic features from CT scans can identify a pancreatic cancer subtype that may respond better to immunotherapy.

## Contribution

The study introduces a radiomics-based approach to non-invasively predict MSI-H/MMRd status in pancreatic cancer patients.

## Key findings

- A radiomics model using tumor features achieved an AUC of 0.73 for predicting MSI-H/MMRd status.
- Combining tumor and pancreas radiomic features improved model performance with an AUC of 0.83.
- The model showed high negative predictive value (95%), suggesting potential for ruling out the subtype.

## Abstract

Pancreatic cancer, mostly presenting as pancreatic ductal adenocarcinoma (PDAC), has a poor prognosis. The microsatellite-instable (MSI-H)/mismatch repair deficient (MMRd) subtype, however, is more susceptible to immune therapy and is expected to have a better prognosis. Presently, MSI-testing is not routinely performed on PDAC. We assessed whether quantitative imaging features (radiomics) of pretreatment computed tomography (CT) scans could diagnose MSI-H/MMRd.

For this pilot study, we analyzed CT-scans of treatment-naïve sporadic or Lynch syndrome (LS)-associated MSI-H or MMRd PDACs, diagnosed or treated in a single center from 2007 to August 2022. CT-scans of resected MSI-stable, MMR proficient, non-LS PDACs formed a control group, after random selection in 1:4 ratio. Upon CT-scan segmentation, 254 well-defined radiomic features were extracted from pancreas and tumor regions. The predictability of the features was assessed within a repeated stratified 3-fold cross-validation framework by designing three models using random forest classifier, with the most discriminating features selected through the minimum redundancy maximum relevance method from three feature sets: tumor radiomics, pancreas radiomics, and combined tumor + pancreas radiomics. Performance was evaluated by area under receiver operating curve (AUC), sensitivity, specificity, positive and negative predictive value.

Overall, 95 patients were included: 19 patients with MSI-H/MMRd/LS (36.8% female; median age at diagnosis 72 [IQR 60–77 years]) and 76 matched controls with PDAC (53.9% female; median age at diagnosis 66 [IQR 57–74 years]). Median year when CT-scan was done was 2017 and 2018, respectively. The model using radiomic features from the pancreatic tumor reflecting MSI-H/MMRd, had an area under receiver operating curve (AUC) of 0.73. The performance of the model was improved by also incorporating radiomic features from pancreas texture (AUC of combined model 0.83 sensitivity 84%, specificity 78%, negative predictive value 95%).

This pilot study suggests that radiomic features could be used to determine MSI/MMRd status in CT-scans of PDAC, but needs further independent multi-site validation in larger cohorts. Routine application of radiomics to determine MSI-status might be of interest in clinical practice to select patients who could benefit from immune therapy.