Masteronderzoek afdeling Urologie, Erasmus MC Kanker Instituut

Title: AI PRECISE — Improving diagnosis, prognosis, and treatment stratification in high-risk non-muscle-invasive bladder cancer using computational pathology (H&E)

Background
High-risk non-muscle-invasive bladder cancer (HR-NMIBC) accounts for a clinically important subgroup of bladder cancer patients with substantial risk of recurrence and progression. Standard of care typically includes transurethral resection followed by intravesical BCG. However, a considerable fraction of patients (often cited ~30–50%) do not benefit from BCG, motivating better pre-treatment stratification and more reproducible pathology assessment.
Pathological evaluation and current risk stratification can suffer from intra/inter-observer variability. Computational pathology offers a route to objective, scalable and clinically deployable decision support by extracting morphological signatures from routine H&E whole-slide images (WSIs).
AI PRECISE (NWO-funded) specifically targets unmet needs in HR-NMIBC by (i) refining molecular subtyping with attention to intratumor heterogeneity, (ii) developing AI models that predict key biology directly from histology as a faster, more cost-effective alternative to RNA-based profiling (BRS classification), and (iii) enabling discovery of actionable targets with downstream functional validation.

Aim of the research project
Track A — AI for prediction from H&E (core computational track)

• Build/benchmark deep learning pipelines on WSIs (tile extraction, QC, stain variability handling, MIL/transformer baselines).
• Predict outcomes relevant to HR-NMIBC (e.g., response proxies and/or recurrence endpoints depending on available labels).
• Optional: integrate clinicopathological covariates and evaluate calibration/clinical utility.

Track B — Efficient labeling & data curation (high-impact, practical track)

• Develop guidelines and workflows to label WSIs more efficiently (annotation strategy, inter-annotator agreement, label schema).
• Create “annotation-to-training” loops to quantify how labeling choices affect model performance.

Track C — Subtype/biology signal discovery from H&E (research track)

• Train models to predict molecular subtype–related signatures from H&E and localize predictive regions (attention maps / concept bottlenecks / spatial transcriptomics).
• Optional: compare learned regions with heterogeneity concepts motivated by AI PRECISE.

Type of research / research activities for the student

• Work with digitized pathology WSIs (H&E) and associated metadata under Erasmus MC governance.
• Produce a reproducible pipeline (documentation + code) and a structured evaluation (cross-validation strategy, external validation if feasible).
• Present progress in weekly group meetings and contribute to internal reporting / manuscript figures where appropriate.

What do we offer?

• Digital pathology fundamentals (WSIs, scanning artifacts, tiling, QC, annotation standards).
• Practical ML for computational pathology (MIL, weak supervision, interpretability, robustness).
• Research data management in a clinical research context (privacy-aware workflows, dataset versioning, audit trails).
• Communication and teamwork in an international translational research group.

Starting moment
June/July or Sept/Oct

Contact details for the student / daily supervisor
Dr. Tongjie Wang (t.j.wang@erasmusmc.nl), Postdoctoral Researcher (Biomedical AI), Erasmus Medical Center

Formal supervisor
Dr. Tahlita Zuiverloon MD PhD, Associate Professor, Erasmus Medical Center

References
de Jong FC, Laajala TD, Hoedemaeker RF, Jordan KR, van der Made ACJ, Boevé ER, van der Schoot DKE, Nieuwkamer B, Janssen EAM, Mahmoudi T, Boormans JL, Theodorescu D, Costello JC, Zuiverloon TCM. Non-muscle-invasive bladder cancer molecular subtypes predict differential response to intravesical Bacillus Calmette-Guérin. Sci Transl Med. 2023 May 24;15(697):eabn4118. doi: 10.1126/scitranslmed.abn4118. Epub 2023 May 24. PMID: 37224225; PMCID: PMC10572776.