Research
Our main goal is to develop cutting-edge AI methods and an appropriate environment that will maximally support cooperation between domain experts and computer science specialists with a focus on explaining the behavior of these AI methods (explainable AI, XAI). As an indispensable part of such effort, we consider traceable development, training, and validation of the AI methods by the means of automated provenance information generation. For specific domains, we will also develop a trusted environment for the validation of AI methods using independent previously unseen data sets. We also strive to operate the system in production execution in specific domains.
Research directions concerning inter-domain communication:
- Development of self-explaining AI methods.
- Development of XAI methods providing insight into the generic AI methods that would allow “communication” of domain experts with the AI systems.
- Development of federated XAI methods.
- Advanced visualization methods including the development of UI usable by specialists in medicine and natural sciences (MDs, pathologists, biochemists).
- Use of explainability to develop novel (X)AI methods by iteratively embedding input from domain experts on explainability results.
Masaryk Memorial Cancer Institute (Dr. Nenutil - Department of Oncological Pathology)
- Data: Extensive source of histopathological imaging (with a production capacity of new data at 1TB/day) and biological samples for further analyses (long term and short term storage), rich structured and unstructured clinical data, genetic data.
- Expertise: Expertise in histopathological imaging - including methods for automated generation of annotations of the material and capacity to develop manual annotations. Expertise in clinical applications with the capability to organize clinical trials.
- Possible collaborations: Cooperation with computer science-aware pathologists who are able to identify computer science problems and even sketch CS solutions. Clinical validation of developed methods.
Medical University Graz (prof. Zatloukal, dr. Müller)
- Data: Massive source of histopathological imaging (e.g., >1PB of just colorectal cancer WSIs), massive archives of biological materials for further analysis, rich structured and unstructured clinical data.
- Expertise: Expertise in histopathological imaging - including methods for automated generation of annotations of the material and capacity to develop manual annotations. Expertise in clinical applications with the capability to organize clinical trials. Human interaction with the AI methods.
- Possible collaborations: Direct collaboration on computer science problems stemming from pathology-related data production and processing. Mostly classical image processing methods, almost no deep learning.
BBMRI-ERIC (assoc. prof. Holub)
- Data: Source of Europe-wide histopathological imaging, clinical data, and genetic data (Europe-wide Colorectal Cancer Cohort with >10,500 patients).
- Expertise: Data harmonization, privacy protection, data provenance (including standardization), data sharing. Access to primary data sources in European hospitals.
- Possible collaborations: Access to readily available data and providing access to the primary data sources when needed.
Medispo (assoc. prof. Matoulek - also 1.LF UK)
- Data: Clinical data (time series of treatment regime, results of treatment, glycemic index, etc.) of donors from bariatry and obesitology.
- Expertise: Expertise in treatment of obese/baryatric patients: design and optimization of individual treatments, monitoring of patients.
- Possible collaboration: Definitions of clinically relevant problems, partner for validation of developed methods (possibly not via Medispo but via 1.LF UK).
MagicWare (Ing. Novosad, de facto SME status)
- Data: Interface to time series of lifestyle data from wearables - Garmin and Apple - for the Medispo donors
- Expertise: Data integration, database hosting, API development, chatbots for monitoring the patients/donors
- Possible collaborations: Access to data from wearables linked directly to the Medispo data. Application of AI methods to the interaction of chatbots with donors/patients to collect medically relevant data in the least intrusive manner and to warn donors/patients about critical conditions
Our team has developed a deep learning system for processing whole-slide pathological images (WSI) which contains the following XAI facilities:
- Methods for detection of various types of cancerous tissue using cutting-edge deep learning techniques. The system is modular and allows rapid prototyping of new deep learning methods for WSI.
- Methods for explanation of input-output behavior of deep neural networks over WSI using various kinds of methods allowing black-box assessment of the networks (saliency maps, occlusions). Our results indicate that best-performing state-of-the-art methods are deciding based on biological features assessed as critical by the expert pathologists. The explainability methods demonstrate their potential to improve the quality of manual annotations (i.e., the explainability methods delineating features more precisely than the original annotations used for training).
- An advanced visualization engine allowing an overview of (X)AI results over WSI. The engine supports direct communication between computer scientists and pathologists, speeds up model diagnostics, and allows data labeling refinement based on results of the above XAI methods.
Our current research concentrates, among other things, on
- Training models combining predictions with explanations based on vision transformers.
- Investigate methods to break the anonymity of the WSI dataset. Our goal is to draw a precise border between completely anonymized data and data allowing deduction of further information about samples/patients.
Capturing provenance information of (X)AI methods training and validation and standardization of provenance models (lead of ISO 23494 standard development).