Evaluating Deep Learning Approaches for Predicting Ki-67 Scores from H&E-Stained Metastatic Prostate Cancer Images

Zhijun Chen, Helen Richards, Haoyue Zhang, Lucas Jing Liu, Erolcan Sayar, Michael Haffner, Stephanie Anne Harmon

Keywords: Ki-67, multiple instance learning, metastatic prostate cancer

TL;DR: Evaluating multiple instance learning for predicting Ki-67 from H&E images

Abstract: Ki-67 is widely used in cancer research as a proliferation marker. Its expression is usually assessed using immunohistochemistry (IHC). Predicting Ki-67 from hematoxylin and eosin (H&E) stained images is difficult since H&E staining does not directly indicate the presence of the Ki-67 protein. However, recent developments in computational pathology and deep learning, especially foundation models and multiple instance learning (MIL), have opened up new possibilities. In this study, we test several MIL methods for predicting Ki-67 scores in a dataset of weakly-labeled metastatic prostate cancer core images. Experiments show that although it is possible to classify cores into high and low Ki-67 groups, accurately predicting Ki-67 scores from H&E images remains a challenging task.

Submission Number: 76