AI-based prognosis and therapy response prediction from pathology images
The principle of platform is as follows:
1. A potent multi-class tissue segmentation algorithm will be applied to hematoxylin&eosin (H&E) stained whole slide images of the tumor and will precisely segment it into all relevant tumor-associated and benign tissue classes (e.g., for tumor regions these are such classes as tumor, tumoral stroma, necrosis, mucin, tertiary lymphoid structures, etc.). This algorithm is specific for each specimen/tumor type (lung, prostate, colorectal, breast cancer)
2. A cell-level algorithm will be applied that detects all single cells and classifies them (tumor cells and 5 types of microenvironment cells).
This platform allows full, quantitative, explainable deciphering of pathology images of malignant tumors.
Based on this information, powerful prognostic (patient survival) and predictive (response to therapy) algorithms can be built. Thousand of different parameters, including combinations based on simple metrics or spatial distribution of tumor and tumor microenvironment structures can be easily established in the platform.
Thus, for all four types of tumors we developed image-based, quantitative, explainable prognostic parameters with independent value when investigated together with common clinicopathological prognostic variables, such as based on tumor-stroma ratio, necrosis and tertiary lymphoid structure quantification, composition of tumor stroma, etc. These parameters work for different endpoints such as progression-free survival, cancer-specific, and overall survival and were tested in own patient cohorts and using one open cohort (The Cancer Genome Atlas).
- Validate the previously developed prognostic image-based parameters (separately for lung cancer, prostate cancer, colorectal cancer, and breast cancer) using pathology images and associated clinical / pathological / follow-up information.
- Develop new prognostic (patient survival) and predictive (therapy response, metastasis development) parameters using large, well-characterized cohorts of patients (PLCO) and above mentioned computational pathology platform.
- Additionally study intratumoral heterogeneity of different image-based parameters as potential confounder of prognostic and predictive information.
- Develop a image-based system for searching of similar cases.
Reinhard Büttner, Institute of Pathology, University Hospital Cologne
Christian Harder, Institute of Pathology, University Hospital Cologne
Zhilong Weng, Institute of Pathology, University Hospital Cologne