ºÚÁϸ£ÀûÍø

Impact of the project

Assessment of PM by conventional radiological imaging is challenging and therefore diagnostic laparoscopy (DLS) is often performed in high-risk patients to assess the presence and extensiveness of PM. Since treatment options for PM are expanding, there is an increasing need to assess presence and extensiveness of PM before, during and after treatment, ideally by non-invasive techniques.[3,4] Compared to DLS, imaging is widely available, less invasive and cheaper, thereby improving healthcare both on social and economic level.[5-8] By enhancing the accuracy of PM assessment on imaging using AI assistance, we expect to improve radiologists’ performance, thereby improving decision-making in terms of treatment initiation, continuation or termination, and minimizing unnecessary procedures, costs and side-effects.

Project summary

Peritoneal metastases (PM) are cancer spread to the peritoneum and are most commonly observed in patients with gastric, colorectal, and ovarian cancer. Accurate assessment of the presence and extent of PM is crucial for determining treatment options and evaluating treatment response. Currently, diagnostic laparoscopy (DLS) is considered the gold standard for assessing PM, but it is an invasive and costly procedure associated with potential risks for the patient. Radiological imaging, such as CT and MRI, is widely used as a less invasive alternative, but accurate detection of PM on imaging is challenging and requires significant expertise. Subtle PM lesions are often missed by non-expert radiologists, while extensive PM is easier to detect. 

This project aims to develop Artificial Intelligence (AI) models to improve and standardize the radiological assessment of PM. Specifically, two AI models will be developed. The first model will automatically divide the abdominal cavity into 13 regions according to the radiological Peritoneal Cancer Index (rPCI), based on consensus from an international panel of experts. This structured approach is expected to reduce interobserver variability and facilitate a more objective assessment of PM. 

The second AI model will focus on the automatic detection and segmentation of PM nodules on CT scans. State-of-the-art deep learning techniques, including Vision Transformers (ViT) and 3D convolutional neural networks, will be applied to recognize subtle patterns associated with PM. The model will be trained on a large retrospective dataset consisting of over 300 anonymized CT scans of patients with and without PM, collected from the Catharina Cancer Institute and referring hospitals. Abdominal regions and PM nodules will be annotated by clinical researchers under supervision of expert radiologists. 

This project is a close collaboration between the Catharina Cancer Institute, Maastricht University, and Eindhoven University of Technology. By leveraging cutting-edge AI techniques, the models are expected to match the performance of expert radiologists in assessing PM on imaging. The ultimate goal is to improve patient selection for treatment, reduce the need for invasive DLS procedures, and enable faster and more accurate diagnosis, especially in non-expert centers. This will contribute to better-informed treatment decisions, reduce healthcare costs, and minimize the physical burden for patients. 

Collaborative partners

• Eindhoven University of Technology (lead)
• Catharina Hospital Eindhoven
• Maastricht University

Funding Programme:

- Main responsible for AI part of project: Dr. ir. F. (Fons) van der Sommen, associate professor computer vision and image processing, PhD, Eindhoven University of Technology
- Main responsible for clinical part of project: Prof. dr. I.H.J.T. (Ignace) De Hingh, gastrointestinal and oncological surgeon, MD PhD, Catharina Cancer Institute and GROW