Abstract

One of the primary causes of death in older men is prostate cancer (PCa), and early detection can lower the death rate. There are numerous methods for making an early diagnosis, including the Digital Rectal Exam (DRE), which involved inserting a gloved hand into the rectum to feel for prostate bumps, the PSA blood test, which measures ng/ml and is used by many doctors to make a rough guess as to whether a patient has PCas (some use 4ng/ml or higher while others can go as low as 2.5ng/ml), among many other high-tech methods. In this effort, a modified deep learning technique employing augmented data and transfer learning was developed for the categorization of the relevance of two (2) lesion types. This was accomplished using a collection of 326 MRI scans of patients who were suspected of having prostate cancer, combined with information about their actual diagnosis and possible tumor location (s). With an accuracy of 81% and an AUC score of 0.79, CNN binary classification was successfully trained, tested, and compared against other models. When fed with rich datasets, this crucial transfer learning can be utilized as an automated decision-making tool to close misdiagnosis gaps.

Key words: Prostrate Cancer detection, deep learning, Machine learning, VGG16.