Robust Learning Methods for Whole Slide Image Analysis in Digital Pathology

​ Deep learning methods are widely used in digital pathology for the analysis of whole slide images (WSIs), supporting tasks such as tissue detection, cancer classification, grading, and segmentation. Existing approaches include convolutional neural networks, multiple instance learning (MIL) frameworks, transformer-based models, and various hybrid architectures. These methods often report strong performance on curated benchmark datasets. In practice, however, pathology data is affected by a range of artefacts introduced during tissue preparation, staining, scanning, and compression. Common examples include staining variability, blur, tissue folds, scanning noise, partial tissue loss, and other acquisition-related imperfections. Such artefacts can substantially affect model predictions and limit their reliability in clinical or research settings. Although artefact robustness has received increasing attention in recent literature, evaluations are often fragmented, task-specific, or difficult to reproduce, and architectural choices are rarely analyzed in a systematic manner. This project focuses on a structured empirical study of how modern WSI analysis models behave under realistic artefacts, combined with the development of a modest but original methodological improvement informed by these findings. Responsibilities During the course of the project, the student will typically be involved in the following activities: Reviewing recent literature on deep learning methods for whole slide image analysis, with a focus on robustness, artefact handling, and quality control. Installing, adapting, and running existing deep learning models using publicly available implementations. Working with whole slide image datasets and designing experimental setups for model evaluation. Applying and analyzing the impact of different artefacts, either through controlled simulation or by using real artefact-affected data. Comparing model behavior across architectures and training strategies using quantitative metrics and qualitative inspection. Documenting experiments carefully to ensure reproducibility. Exploring potential improvements or extensions to existing methods based on observed limitations. Discussing results regularly with supervisors and adapting the project direction as needed. What You Will Learn Through this project, the student will gain practical and conceptual experience in several areas, including: Hands-on experience with modern deep learning architectures used in digital pathology. Practical skills for working with whole slide images, large-scale data, and GPU-based training pipelines. Understanding how real-world artefacts affect model performance and reliability. Designing and conducting systematic, reproducible experiments in applied machine learning research. Critical evaluation of research papers beyond reported benchmark scores. Translating empirical observations into concrete methodological ideas. Communicating technical findings clearly through written reports, figures, and presentations. Managing an open-ended research project and making informed decisions when initial assumptions do not hold. The project will be conducted as a research internship at the VUB campus in Ixelles, Brussels. Depending on interest and progress, the work carried out during the internship may be further extended into a Master’s thesis, allowing continuity within the same research topic, subject to discussion and approval by the supervisors. Type of internship : Master internship Duration : 6-9 Required educational background : Biomedical engineering, Bioscience Engineering, Computer Science Supervising scientist(s) : For further information or for application, please contact Saeed Mahmoudpour ( Mahmoudpour.Saeed@imec.be ) and Tien Nguyen ( Tien.Nguyen@imec.be ) The reference code for this position is 2026-INT-007 . Mention this reference code in your application. Applications should include the following information: resume motivation current study Incomplete applications will not be considered