Quantitative analysis of tissue morphology and biomarkers with convolutional neural networks for improve prognostics of prostate cancer
Appropriate and timely treatment of prostate cancer is the most important factor determining disease outcome for patients. However, treatment selection is hampered by subjective qualitative grading through histopathological analysis of tissue by pathologists. Large scale introduction of whole-slide digital pathology has opened the door for computerized analysis of histopathological slides. The aim of this project is to improve prognosis estimation and as such treatment planning for patients by turning subjective qualitative grading into reproducible quantitative grading using innovative deep learning techniques.
Deep learning is a technique that is currently mostly used in generic computer vision tasks like image classification. In contrast to traditional image analysis algorithms often applied in medical imaging, deep learning does not require engineers to manually design discriminative features to be extracted from the data. Instead, deep learning systems learn the features which are most discriminative directly from the data itself.
Based on the digitized PLCO prostate whole-slide images and tissue micro-arrays we will train deep convolutional neural networks to directly predict the patient outcome with respect to biochemical recurrence and prostate cancer-related death based on H&E-stained cancer cores. This will allow us to specifically characterize the morphological features which are relevant for assessing prognosis. The prognostic features obtained by this automated system will be compared to the standard Gleason grading of the tissue performed by pathologists via Cox proportional hazard modelling. This allows us to asses whether computationally obtained features could potentially serve as an addition to manual human grading in assessing prostate cancer prognostics. In addition, we will evaluate an existing algorithm based on convolutional neural networks that can fully automatically predict the Gleason grade of this specimen relative to manual grading on this dataset (https://arxiv.org/abs/1907.07980).
- Assess the predictive power of an existing automated Gleason grading system in terms of Cox proportional hazard modelling to predict biochemical recurrence
- Compare this predictive power to manual Gleason grading by experienced pathologists
- Directly predict biochemical recurrence based on tumor tissue morphology without explicitely predicting Gleason grades
- Compare this direct model to the automatic and manual Gleason grading models
Wouter Bulten, MSc, Radboud University Medical Center, Nijmegen, the Netherlands
Jeroen van der Laak, Radboud University Medical Center, Nijmegen, the Netherlands
Christina Hulsbergen - van de Kaa, PhD, LabPON, Hengelo, the Netherlands