KTI

KTI

2021-2023 – funded by the Ministry for Social Affairs, Health and Integration of Baden-Württemberg

The ultimate goal of the AI translation initiative (KTI) project is to create the base for a successful introduction of diagnostic AI based image analysis algorithms into the daily clinical practice by improving their explainability (XAI) as well as generalisation.

The current focus of the project is on pathological diagnostics.

Initially, the XAI methods are evaluated from a technical point of view to determine whether they accurately and reproducibly reflect and elucidate functioning and decision-making of the diagnostic algorithms. The primary aspects to be analyzed and compared are correctness and robustness. Potential users will be interviewed to evaluate the influence of the explainability techniques on acceptance of AI-based assistance systems. Additionally, the direct usefulness of such systems regarding diagnostic precision will be determined.

For the clinical application of KI based assistance systems, one question is of particular importance: In which cases can the practitioner trust the AI-based classification and in which cases can he or she not? For this particular decision, information on the confidence level of the system’s decision might be rather important for the user of a KI based assistance system and might help to estimate the limitations of the system. If the confidence value for a certain image is low, then the training of the model was not sufficient for a well-informed prediction. This could be due to the particular training method that was used and/or to the image being an outlier that is not sufficiently represented in the training data.

Another central requirement for a successful introduction of KI based diagnostic assistance systems into clinical practice is a reliable applicability also an unknown, ‘foreign’ data. Especially in the KI based analysis of histological tissue sections, even small differences in the staining protocol between labs represent a huge challenge. In order to increase the generalization ability of algorithms, different methods (such as normalization and augmentation techniques as well as few shot learning approaches) are tested.