Artificial Intelligence for Prostate Cancer Screening and Diagnosis: A State-of-the-art Review

Abstract

Prostate cancer (PCa) is one of the most prevalent forms of cancer affecting men worldwide. This thesis conducts a state-of-the-art review of AI-based approaches for PCa screening and diagnosis, focusing on three major types of input data: prostate-specific antigen (PSA) screening data, magnetic resonance images (MRIs), and histopathology images. The purpose of this thesis is to determine the current extent of research into AI for PCa screening and diagnosis, highlighting recent advancements, future prospects, and remaining challenges.

The selected articles show wide variability in study designs, data sources, model architectures, and evaluation approaches. Key findings suggest that even though recent research mostly focuses on AI for medical image analysis, PSA data still carries valuable information and can be combined with MRI data and other clinical variables to improve diagnostic performance. AI models inputting MRIs have demonstrated diagnostic performance surpassing the PI-RADS (prostate imaging-reporting and data system) on several coarse-level PCa classification tasks. Similarly, AI models employed for Gleason grading of histopathology images have been reported to match pathologists' performance and to improve inter-rater agreement.

In addition to the literature review, an analysis consisting of 11 articles was conducted to identify the relationship between the AUC values, cohort sizes, and model types of models differentiating between clinically significant PCa (csPCa) and non-csPCa using MRIs. The analysis reveals a statistically significant positive correlation (Kendall's tau = 0.673, 95% CI: 0.348-0.944, p = 0.020) between the AUCs and the cohort sizes, suggesting that increasing cohort size might have a positive impact on performance. In addition, permutation tests with the t-test statistic indicate a statistically significant difference (t = 2.297, p = 0.032) between the mean AUC values of traditional machine learning models and deep learning models, suggesting that deep learning models might be capable of achieving better performance than traditional machine learning models on such medical image analysis tasks.

Despite the great potential and significant progress of research into AI for PCa, several challenges remain, including the shortage of data, label noise, and wide variability in study implementation and evaluation. These problems call for the application of novel machine-learning techniques and collaborative research endeavors.