Survival prediction and extraction of novel microscopic diagnostic patterns in prostate cancer
Principal Investigator
Name
Johannes Lotz
Degrees
Ph.D.
Institution
Fraunhofer-Gesellschaft zur Förderung der angewandten Forschung e.V.
Position Title
Principal Scientist
Email
About this CDAS Project
Study
PLCO
(Learn more about this study)
Project ID
PLCOI-1612
Initial CDAS Request Approval
Jul 8, 2024
Title
Survival prediction and extraction of novel microscopic diagnostic patterns in prostate cancer
Summary
The project aims to predict the survival of patients with prostate cancer using deep learning based solely on patterns in hematoxylin-eosin (H&E) stains from prostatectomy specimens. The following questions are addressed:
1. Can biochemical recurrence and/or survival be predicted based on H&E-sections before or at the time of prostatectomy?
2. Can morphological patterns that are responsible for shorter survival be identified?
We are currently developing these methods using internal data. We are preparing the release of a multi-center prostate cancer survival dataset from the university hospitals of Berlin (Charité) and Frankfurt. External data is required to validate the developed methods. The PLCO dataset would serve as an external validation dataset. Additionally, we will extend the scope of the project to determine whether:
3. similar methods can be used to predict therapy response in these patients and
4. these methods can be transferred to other entities, such as lung, colorectal and ovarian cancers.
Methods
We combine pre-trained foundation models and predictive pattern recognition a) to predict the recurrence-free survival of individual patients and (if possible) response to therapy and b) infer the relevant features in the tissue morphology to identify outcome-determining structures. We perform the training of the deep learning pipeline in two steps:
Step 1, extract predictive patterns: The WSI data will be analyzed and condensed to clinically predictive patterns using a foundation model that has been pre-trained specifically for prostate tissue (https://doi.org/10.48550/arXiv.2311.09847, accepted at Nat. Comput. Sci., https://github.com/FraunhoferMEVIS/MedicalMultitaskModeling). This step reduces the image data to data-driven features similar to concepts such as the different morphological types of cell nuclei. By extracting patterns across multiple datasets, a novel abstract representation is learned that is less dependent on the particularities of a single clinical laboratory site, such as staining or preparation bias. The aim of this step is to improve the cross-center robustness of the resulting model.
Step 2, train prediction of outcome: Based on the representations, a predictor for survival and/or therapy response is trained. Once successful, the most relevant representations are clustered and linked back to the respective tissue patches to correlate them with physiologically relevant patterns.
This approach allows for an explainable analysis of prostate tissue sections. If successful, the outcome can be linked to a combination of physiological features beyond ISUP scores, and new morphological biomarkers can be developed.
All results will be published in open access journals, and developed software will be made available to the research community.
1. Can biochemical recurrence and/or survival be predicted based on H&E-sections before or at the time of prostatectomy?
2. Can morphological patterns that are responsible for shorter survival be identified?
We are currently developing these methods using internal data. We are preparing the release of a multi-center prostate cancer survival dataset from the university hospitals of Berlin (Charité) and Frankfurt. External data is required to validate the developed methods. The PLCO dataset would serve as an external validation dataset. Additionally, we will extend the scope of the project to determine whether:
3. similar methods can be used to predict therapy response in these patients and
4. these methods can be transferred to other entities, such as lung, colorectal and ovarian cancers.
Methods
We combine pre-trained foundation models and predictive pattern recognition a) to predict the recurrence-free survival of individual patients and (if possible) response to therapy and b) infer the relevant features in the tissue morphology to identify outcome-determining structures. We perform the training of the deep learning pipeline in two steps:
Step 1, extract predictive patterns: The WSI data will be analyzed and condensed to clinically predictive patterns using a foundation model that has been pre-trained specifically for prostate tissue (https://doi.org/10.48550/arXiv.2311.09847, accepted at Nat. Comput. Sci., https://github.com/FraunhoferMEVIS/MedicalMultitaskModeling). This step reduces the image data to data-driven features similar to concepts such as the different morphological types of cell nuclei. By extracting patterns across multiple datasets, a novel abstract representation is learned that is less dependent on the particularities of a single clinical laboratory site, such as staining or preparation bias. The aim of this step is to improve the cross-center robustness of the resulting model.
Step 2, train prediction of outcome: Based on the representations, a predictor for survival and/or therapy response is trained. Once successful, the most relevant representations are clustered and linked back to the respective tissue patches to correlate them with physiologically relevant patterns.
This approach allows for an explainable analysis of prostate tissue sections. If successful, the outcome can be linked to a combination of physiological features beyond ISUP scores, and new morphological biomarkers can be developed.
All results will be published in open access journals, and developed software will be made available to the research community.
Aims
- Validate the prediction of biochemical recurrence and/or survival based on H&E-sections before or at the time of prostatectomy.
- Discovery and validation of morphological patterns that are responsible for shorter survival be identified.
- Evaluate if therapy response can be predicted based on similar patterns.
- Evaluate if the identified patterns can be transferred to other entities, such as lung, colorectal, breast, and ovarian cancers.
Collaborators
Tim-Rasmus Kiehl, Institute of Medical Informatics, Charité – Universitätsmedizin Berlin, Germany
Sefer Elezkurtaj, Institute of Pathology, Charité – Universitätsmedizin Berlin, Germany
Peter Wild, Senckenberg Institute of Pathology, University Hospital Frankfurt, Germany