Automating Analysis of Biomarkers Ki-67 and PSA Using Deep Learning in Prostate Cancer

Early detection and diagnosis of prostate cancer play an important role in improving patient outcomes. Two biomarkers, Ki-67 and prostate specific antigen (PSA), can provide insight into tumor aggressiveness. Currently, pathologists manually analyze biopsy scans, which can be time-consuming and subjective, and therefore delay the results. This thesis explores the implementation of an automated pipeline to efficiently scale image preprocessing, generating data suitable for learning-based analysis. Whole-slide images of prostate tissue stained for Ki-67, PSA, and hematoxylin and eosin (H&E) were preprocessed. An automated pipeline was developed using modular scripts to make it reproducible. The pipeline streamlines all preprocessing steps with minimal manual intervention. The work highlights the importance of image preprocessing and workflow automation in order to prepare data for machine learning models. From the results, we can see that the pipeline produces the same results when running on different operating systems. Furthermore, the patches created for each biopsy correspond to the same coordinates, which means that the patches can be used for machine learning inference.