Skip to Content

Deep Learning Methods for Automated Gleason Grading

Team Members:
  • Yujie Zhao
  • Ananya Tandri
  • Ruitao Hu
  • Yuxin Du
  • Zhenzhen Wang
  • Adam Charles


Prostate cancer remains the second most common cause of death among men worldwide. It is challenging to treat, with early detection not always associated with improved outcomes. In diagnosing prostate cancer, the Gleason Grading System is an invaluable tool. Abnormal prostate cells are assigned Gleason scores, which range from one to five, where grade 1 signifies normal prostate cells and grade 5 suggests significant abnormalities in these cells. Yet, there are significant challenges in creating consistent, objective scores. Even among experienced pathologists, Gleason grading remains somewhat subjective. Deep learning-based image analysis provides a promising route to automate the production of Gleason grades, yet current models have deficiencies in quality and generalizability. Here, we describe a Histopathology-Pretrained-Based CNN to address these issues, providing additional focus to the discrepancies that may naturally arise in biopsy processing, such as color differences from scanners and whole mount vs biopsy analysis.

Read the Johns Hopkins University privacy statement here.