Scientific collaboration in BioMedAI was significantly strengthened, leading to multiple proposal submissions and several funded follow-ups, including RI-SCALE (Horizon Europe), an AZV project with IKEM, and a GAČR project. A concrete result was the cross-institutional MU–MUG XAI development team focused on FAIRification and management of large digital pathology collections and on developing AI models for medical applications. Collaboration expanded beyond the consortium (e.g. Faculty Hospital St. Anna Brno, IKEM Prague). At MU, the affiliated RationAI laboratory expanded and introduced a clearer leadership structure for efficient mentoring and talent development. The project also catalysed the creation of the AigoPath startup and strengthened industry partnerships (Magicware, Carebot, Comprimato) and strategic collaboration with EMPAIA (via TUB), increasing visibility and access to international networks.
A major practical output is the Educational XAI Toolkit: RatioPath (libraries, tools and templates for digital pathology ML/XAI), RatioVis (interactive experiment reporting and visualization over whole-slide images via the xOpat WSI viewer and an MLflow dashboard), and RatioCast (AI time-series forecasting integrated with SensitiveCloud and tested on blood biomarker data). The toolkit was applied in multiple digital pathology use cases, including prostate cancer (detection and Gleason scoring), colorectal cancer, and breast cancer (Ki-67 scoring). Input data quality control proved critical; AI and non-AI QC/QA methods were implemented as pre-training gates and integrated into the pipelines.
Secure infrastructure for sensitive data was established through migration to the ISO 27001-certified, Kubernetes-based SensitiveCloud at MU (CERIT-SC) and porting RatioPath/RatioVis accordingly. Data management mechanisms ensured strict separation of institutional datasets, controlled access, auditability, and reproducible workflows for pseudonymized WSI used in infrastructure development, pipeline testing, and partner training. A privacy risk model was introduced to systematically assess and mitigate disclosure and re-identification risks and to guide technical/organizational controls. The Data Management Plan (GDPR, FAIR, FAIR-Health) was updated twice and informed a future-proof concept for integrated management of clinical, histopathological and imaging data.
Training and knowledge transfer were delivered through two Summer Schools on explainable AI (expert talks plus extensive hands-on sessions) and a joint Brno–Graz Winter School extending XAI with NLP/LLM methods for structured extraction from clinical datasets. Additional workshops and trainings covered data generation/extraction, annotation workflows, and QA for both pathologists and AI researchers. Junior pathologists and students were trained to review and annotate data, enabling MMCI testing of prostate cancer prototype diagnostic support. Continuous training in FAIR/FAIR-Health data management and sensitive data handling supported the jointly developed tissue classification system, which underpins anonymized digital pathology archives intended for sharing via BBMRI.
