Authors:
Markus Bauer
1
;
Lennart Schneider
2
;
Marit Bernhardt
2
;
Christoph Augenstein
1
;
Glen Kristiansen
2
and
Bogdan Franczyk
3
;
4
Affiliations:
1
ScaDS.AI, University of Leipzig, Germany
;
2
Institute of Pathology, University Hospital Bonn, Germany
;
3
University of Leipzig, Germany
;
4
Wroclaw University of Economics, Poland
Keyword(s):
Artificial Intelligence, Vision Transformer, Self-Supervised Learning, Digital Pathology, Prostate Carcinoma.
Abstract:
The histopathological analysis of prostate tissue is challenging due to the required expertise and the inherently high number of samples. This accounts especially for prostate cancer (PCa) assessment (tumour grading), as parameters like the Gleason score have high prognostic relevance, but suffer from significant interobserver variability, mainly due to individual grading practice and experience. AI-based solutions could assist pathological workflows, but their integration into clinical practice is still hampered, as they’re optimised based on general AI-metrics, rather than clinical relevance and applicability. Moreover, commercial solutions often provide similar performance than academic approaches, are expensive, and lack flexibility to adapt to new use cases. We investigate the requirements to provide a flexible AI-based histopathological tissue analysis tool, that makes the expertise of experienced pathologists accessible to every hospital in a user-friendly, open-source solutio
n. The proposed software allows for slide inspection, tumour localisation and tissue metric extraction, while adapting to different use cases using a Python-enabled architecture. We demonstrate the value of our tool in an in-depth evaluation of transurethral hyperplastic resection tissue (TURP)-chip analysis and PCa grading using a set of extensively annotated prostate cancer patient cases. Our solution can support pathologists in challenging cases, fasten routine tasks and creates space for detail analysis.
(More)