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

Prostate 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.

learning (ML) with neural networks, have been developed to make an accurate diagnosis. 27 In PCa diagnosis, there have been many developed models of ML and the model 28 algorithm using ultrasound images shows good accuracy. This study aims to analyse the 29 accuracy of neural network machine learning models in prostate cancer diagnosis using 30 ultrasound images. The protocol was registered with PROSPERO registration number 31 CRD42021277309. Three reviewers independently conduct a literature search in five 32 online databases (MEDLINE, EBSCO, Proquest, Sciencedirect, and Scopus). We 33 screened a total of 132 titles and abstracts that meet our inclusion and exclusion criteria. 34 We included articles published in English, using human subjects, using neural networks 35 machine learning models, and using prostate biopsy as a standard diagnosis. Non 36 relevant studies and review articles were excluded. After screening, we found six articles 37 relevant to our study. Risk of bias analysis was conducted using QUADAS-2 tool. Of the limited and low-moderate quality evidence, we managed to analyse the predominant . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted February 5, 2022. ;https://doi.org/10.1101https://doi.org/10. /2022  Ultrasound is a potential candidate for PCa imaging because it is cost-effective, practical, 65 and widely available. The problem of ultrasound images interpretation is that hypoechoic 66 areas suspected of cancer can be normal or cancerous histologically. Most prostate . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted February 5, 2022.

2021
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted February 5, 2022. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted February 5, 2022. ; https://doi.org/10.1101/2022.02.03.22270377 doi: medRxiv preprint 8 126 127 We included all available articles about machine learning use in prostate cancer diagnosis 128 that used ultrasound images. We restricted our search to articles published in English 129 without a publication date limit. A study was considered relevant if it fulfilled our inclusion 130 criteria: using human subjects, using neural networks machine learning models, and 131 using prostate biopsy as a standard diagnosis. We included cohort, case-control, and CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted February 5, 2022.

150
Our electronic search identified 145 articles and only 6 that met our inclusion and 151 exclusion criteria (Fig 1). Four articles use the artificial neural networks (ANN) method, 152 one article uses the recurrent neural network (RNN) method, and one article uses the 153 deep learning (DL) method. The characteristics of each study are described in Table 2. 154 The six included studies used a cross-sectional study design. All articles studied adult 155 male human subjects with an unknown age range due to the unclear data. The sample 156 size ranges from 61 to 1077 patients; however, a study from Ronco et al.  (Table   174 2). However, a study from Loch et al. 25 used percentage only. The performance results 175 can be seen in Table 2. Due to the varied parameters, a quantitative analysis could not 176 be performed. Most of the articles used ROC-AUC as the accuracy parameters.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted February 5, 2022.  . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted February 5, 2022. ML can generate input data from various variables to classify whether the patient is 245 suspected of having prostate cancer or not (Fig 3). is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted February 5, 2022.  313 We provide all available evidence about machine learning models of human ultrasound 314 images in prostate cancer diagnosis. However, none of the articles shows the same 315 output parameters to generate a quantitative analysis. Our approach included a 316 comprehensive search of multiple databases as well as other sources for relevant 317 publications. Since we restricted our literature search in English, some articles in other 318 languages may be missed out. The major weakness of this study is low to moderate 319 quality of included studies and the limited number of studies. Although there was only 320 limited evidence, we managed to analyse the predominant findings comprehensively.

Limitation of the Study
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted February 5, 2022. Machine learning with neural network models is a potential technology in prostate cancer 324 that could provide instant information for further workup with relatively high accuracy 325 above 70% of sensitivity/specificity and above 0.5 of ROC-AUC value. Image-based 326 machine learning models would be helpful for doctors to decide whether or not to perform 327 a prostate biopsy. Future development of this technology will be further beneficial in 328 making a diagnosis and treatment evaluation and patient prognosis.