The accuracy of machine learning models using ultrasound images in prostate cancer diagnosis: A systematic review

Abstract

ABSTRACTProstate Cancer (PCa) is the third most commonly diagnosed cancer worldwide, and its diagnosis requires many medical examinations, including imaging. Ultrasound offers a practical and cost-effective method for prostate imaging due to its real-time availability at the bedside. Nowadays, various Artificial Intelligence (AI) models, including Machine learning (ML) with neural networks, have been developed to make an accurate diagnosis. In PCa diagnosis, there have been many developed models of ML and the model algorithm using ultrasound images shows good accuracy. This study aims to analyse the accuracy of neural network machine learning models in prostate cancer diagnosis using ultrasound images. The protocol was registered with PROSPERO registration number CRD42021277309. Three reviewers independently conduct a literature search in five online databases (MEDLINE, EBSCO, Proquest, Sciencedirect, and Scopus). We screened a total of 132 titles and abstracts that meet our inclusion and exclusion criteria. We included articles published in English, using human subjects, using neural networks machine learning models, and using prostate biopsy as a standard diagnosis. Non relevant studies and review articles were excluded. After screening, we found six articles relevant to our study. Risk of bias analysis was conducted using QUADAS-2 tool. Of the six articles, four articles used Artificial Neural Network (ANN), one article used Recurrent Neural Network (RNN), and one article used Deep Learning (DL). All articles suggest a positive result of ultrasound in the diagnosis of prostate cancer with a varied ROC curve of 0.76-0.98. Several factors affect AI accuracy, including the model of AI, mode and type of transrectal sonography, Gleason grading, and PSA level. Although there was only limited and low-moderate quality evidence, we managed to analyse the predominant findings comprehensively. In conclusion, machine learning with neural network models is a potential technology in prostate cancer diagnosis that could provide instant information for further workup with relatively high accuracy above 70% of sensitivity/specificity and above 0.5 of ROC-AUC value. Image-based machine learning models would be helpful for doctors to decide whether or not to perform a prostate biopsy.