Large-scale Whole Slide Image Analysis with Deep Learning Methods to Improve Prostate Cancer Diagnosis
Gleason grading system serves as an essential component in risk stratification and treatment planning for prostate cancer patients. Currently, Gleason grading relies on pathologists to examine glass tissue slides at scanning magnification and localize suspicious regions for higher power examination. Such process can be time-consuming and prone to inter- and intra- observer variability. Moreover, the Gleason grading system may be constrained by its categorization system, which cannot fully capture the disease's continuous feature spectrum. With the recent development of digital slide scanners and the approval of using digitized slides for primary diagnosis by the Food and Drug Administrations (FDA), large numbers of traditional glass slides have been digitized into high resolution whole slide images (WSIs), opening opportunities in developing computational image analysis tools to reduce pathologists' workload and potentially improve inter- and intra- observer agreement.
This research attempts to address several challenges in WSI analysis and assist the histopathologic evaluation of prostate cancer. A tile-level semantic segmentation model, which explicitly leverages multi-scale representations, is first proposed to generate pixel-wise Gleason pattern predictions and facilitate estimation of percentage of different patterns. An expectation-maximization (EM)-based semi-supervised learning framework is then developed to exploit information embedded in weakly-labeled samples to further improve the performance of segmentation models, which alleviates the need of expensive pixel-wise annotations. Besides building segmentation tools for tiles extracted from WSIs, a novel multi-resolution multiple instance learning-based model, which can be trained with slide-level labels, is proposed to identify informative regions and provide slide-level Gleason grade group prediction. The model is developed and validated on a large-scale prostate biopsy dataset. Furthermore, a deep learning system, which leverages histopathological features and attention-based aggregation, is built to facilitate predictions of progression-free survival after radical prostatectomy. Together, these models demonstrate the potential of several computer aided diagnosis tools, and pave the road for utilizing computational approaches to optimize and improve the efficiency of prostate cancer diagnosis and risk stratification.