Bart de Keizer

UMC Utrecht

In short, Bart de Keizer's research entails the following: 

Artificial intelligence-based phenotyping of metastatic renal cell carcinoma to better understand (pseudo)progression under immune checkpoint inhibition

This project aims to determine if there are differences in the way that metastatic renal cell carcinoma (cancer originating from the kidneys that has spread to lymph nodes and/or other organs) appears on PET/CT imaging scans between cases where the cancer has truly progressed and cases where it appears to have progressed but is actually responding to treatment (called pseudo progression). If such differences do exist, it may be possible in the future to switch patients from one type of treatment (called immune checkpoint inhibitors) to another type (called targeted therapy) earlier, in order to stop the cancer from progressing further while also reducing unnecessary side effects.

Immune checkpoint inhibitors are often continued when the cancer appears to be growing because of the possibility of pseudo progression, but these treatments can cause severe side effects in some patients. If a way can be found to distinguish between true progression and pseudo progression using imaging scans, it could help doctors decide whether to continue immune checkpoint inhibitors or switch to targeted therapy, potentially prolonging the time before the cancer progresses and reducing side effects.

Currently, the distinction between true progression and pseudo progression is made on the basis of the immune Response Evaluation Criteria in Solid Tumors (iRECIST), which involves taking multiple imaging scans over a period of time to see if the cancer responds to treatment. However, this process can take several weeks or even years.

To find faster and more objective response characteristics early after start of treatment we will use machine learning (deep learning and radiomics), combined with explainable artificial intelligence.

The project a collaboration between the departments of radiology, Image Sciences Institute, and medical oncology at UMC Utrecht. In addition, three other university medical centers participate as well.

 

Short summary of progress/results April 2026

In this project, we use artificial intelligence (AI) to better understand how patients with metastatic renal cell carcinoma respond to immunotherapy. We collected CT scans and clinical data from more than 650 patients across nine hospitals and developed automated methods to detect and track tumor lesions over time.

Our results show that AI can improve the consistency of treatment response assessment between radiologists and enables more detailed analysis of all tumor lesions within a patient. In addition, we are developing AI models that can predict treatment response within six months and identify pseudoprogression. These insights may contribute to more personalized and effective treatment strategies in the future. 

Short summary of progress/results April 2025

In the second year of this study, we maintained data collection from other institutes, which is now almost completed. A total of nine institutes have now joined the study and  we expect to complete our training dataset of approximately 500 patients in the upcoming months. The first goal of this project is to automate response measurements  of metastatic renal cell carcinoma to immunotherapy. This pipeline consists of a segmentation model, for which we trained state-of-the-art networks. Furthermore, we developed a metastasis tracking algorithm that allows us to automatically analyze how these metastases change over time. The preliminary results of this network
are currently being compared to clinical scores.

Short summary of progress/results April 2024

In the first year of this study, a database was constructed for the TRIPP study, including patients with metastatic renal cell carcinoma in UMC Utrecht who received immune checkpoint inhibitors between 2019-2023. For each patient, three longitudinal CT scans are included, along with the necessary clinical information and response measurements. We have currently included 59 patients with 177 scans.

Because this is a multi-center study, multiple agreements and waivers have been written and received to transfer data in a safe manner. Until now, four institutes have agreed to join our study and we expect to collect data in the following months. Additionally, numerous other (academic) medical centers have also been contacted to contribute.

The first part of this project focuses on automating response measurement. The first step in this process is the segmentation of the metastatic lesions in the images. For this, a state-of-the-art convolutional neural network has been trained.