Grading of Prostate Whole-slide Images Using Weak Self-supervised Learning

Amirata Ghorbani, Andre Esteva, James Zou

Grading of Prostate Whole-slide Images Using Weak Self-supervised Learning

Amirata Ghorbani, Andre Esteva, James Zou

Abstract: Prostate cancer is the second most common cancer among men worldwide. For prognosis and treatment, pathologists assign International Society of Urological Pathology (ISUP) grades to Whole Slide Imaging (WSI) of sampled prostate tissues to express the severity of the cancer. The manual approach suffers from human error and person-to-person variations. Existing approaches that use machine learning for automatic grading of WSI images includes using expensive datasets in which different regions of the slide are annotated by pathologists to show different levels of cancer. However, most of the existing real-world datasets contain weak labels; i.e. each slide is labeled with just one grade number. In this work, we present a self-supervised learning method to automatically grade prostate slides using datasets of weakly labeled slides.

Poster

Abstract: Prostate cancer is the second most common cancer among men worldwide. For prognosis and treatment, pathologists assign International Society of Urological Pathology (ISUP) grades to Whole Slide Imaging (WSI) of sampled prostate tissues to express the severity of the cancer. The manual approach suffers from human error and person-to-person variations. Existing approaches that use machine learning for automatic grading of WSI images includes using expensive datasets in which different regions of the slide are annotated by pathologists to show different levels of cancer. However, most of the existing real-world datasets contain weak labels; i.e. each slide is labeled with just one grade number. In this work, we present a self-supervised learning method to automatically grade prostate slides using datasets of weakly labeled slides.