Jeanin van Hooft

LUMC

In short, Jeanin van Hooft's research entails the following:

Improving early detection of PANcreatic cancer in HIgh-risk individuals through Artificial Intelligence methods (PAN-HI-AI)

Pancreatic cancer (PC) has a dismal survival due to most patients presenting with advanced stage disease. Magnetic resonance imaging (MRI) and endoscopic ultrasound (EUS) surveillance programs for high-risk populations, whose lifetime risk of PC varies between 5-35%, have been established.

There is a need to improve the modalities used to surveil these high-risk individuals (HRI), as only a minority of tumors are discovered as precursor lesions with high-grade dysplasia (HGD) or early-stage cancers. Artificial intelligence (AI) applied to MRI, specifically radiomics and deep learning models, could aid by complex pattern recognition of subtle changes associated with progression to malignancy in these HRI, enabling detection at a curable stage as precursor lesions or early-cancers.

A previously developed automated pancreas segmentation algorithm will be validated on the complete LUMC MRI dataset of 2400 longitudinal MRIs from 340 individuals. Subsequent external validation is assured through the 1200 MRI database of Sweden’s Karolinska Institute.

After the segmentation algorithm has delineated the pancreatic region of interest, the LUMC and Karolinska cohorts will be used for training and validating a diagnostic classification algorithm for detecting PC. The classification algorithm will take a longitudinal approach, incorporating multiple MRI sequences and demographic variables into predicting precursor lesions with HGD and/or cancer.

The final step is reserved for a robust clinical trial design aiming to prospectively validate the longitudinal classifier in HRI surveillance cohorts.

 

Short summary of progress/results April 2026

Our study started at the end of 2024 with the aim of detecting early-stage pancreatic cancer (and precursor lesions) earlier on MRI using artificial intelligence. This research is conducted within a cohort of high-risk individuals (approximately 20% cumulative risk due to a genetic mutation), who are followed annually with MRI scans at the LUMC. 

Over the past year, we have further characterized the high-risk cohort by adding clinical and imaging data to our database. In parallel, we are curating the dataset to optimally leverage multiple MRI sequences and are establishing external validation cohorts, along with the required collaboration agreements. In addition, we are working on improving our pancreas segmentation model. This model was originally trained on T2-weighted MRI scans, but we are expanding it to other MRI sequences. In this way, we can optimally leverage the complementary information from different MRI sequences.

Alongside this, we have worked on a ‘slice-to-volume reconstruction’ project, in which we combine two MRI scans into a higher-resolution 3D volume. A typical MRI scan consists of multiple 2D sequences that have high in-plane resolution but lower through-plane resolution due to the slice spacing between MRI slices. By combining an axial T2-weighted MRI (from the upper to the lower abdomen) with a coronal T2-weighted MRI (from the front to the back), we can reconstruct a 3D volume with high resolution in all three directions. Our rationale is that starting from a higher-resolution volume can improve the accuracy of downstream AI applications. This initial pilot study shows promising results and has been submitted to a medical imaging conference. In the coming period, we aim to further validate the accuracy of these reconstructions and evaluate their added value.

In addition, over the next year we aim to focus on the actual detection of pancreatic abnormalities, as a first step toward identifying the subtle early-stage pancreatic tumors that we ultimately aim to detect in this project. Overall, the project is progressing well, and we are optimistic about meeting the project goals within the planned timeline.

Short summary of progress/results April 2025

The aim of this study is to enable earlier detection of early-stage pancreatic cancer and precursor lesions on MRI scans using artificial intelligence. 

In November 2024, the clinical PhD candidate (physician-researcher) joined the project. During the initial phase, the focus was primarily on organizational aspects of the research. We drafted a research protocol and submitted it for approval to the relevant ethics committees. Additionally, we tried to obtain informed consent from as many patients as possible within our pancreatic surveillance program for the use of their MRI scans and associated data in this study. We are currently exporting the imaging data (primarily MRI scans, but also CT scans) of patients who have provided consent from the PACS to our network drive, in file formats suitable for further analysis. These scans will also be uploaded to our XNAT facility. In March 2025, the technical PhD candidate also joined the project, completing our research team. This has allowed us to take significant first steps in our research. Using a continuous learning approach, we have developed an initial version of our pancreas segmentation algorithm, which is yielding promising early results. We are currently preparing our first scientific publication on these findings. In the coming period, we will further optimize the segmentation algorithm by expanding the training dataset. We then plan to compare its performance to publicly available pancreas segmentation algorithms and evaluate its performance on MRI scans from external cohorts. 

In parallel, we are working on improving MRI resolution by combining different scan planes (axial, coronal, sagittal) using a super-resolution approach.