APPLICATION OF DEEP LEARNING FOR AUTOMATED PROSTATE CONTOURING
ON MAGNETIC RESONANCE IMAGING DATA: PRELIMINARY RESULTS
Talishinskii A.E.1,2,3,4, Kamyshanskaya I.G.1,2,5, Kryuchkova O.V.6,
Bakhtiozin R.F.7, Pashkovskaya A.A.8
1 - St. Petersburg State University. Saint Petersburg, Russia.
2 - Med-Rey LLC. Moscow, Russia.
3 - Astana Medical University. Astana, Kazakhstan.
4 - Fergana Medical Institute of Public Health. Fergana, Uzbekistan.
5 - Mariinsky City Hospital. Saint Petersburg, Russia.
6 - Central Clinical Hospital with Polyclinic of the Administrative Department of the President of the Russian Federation. Moscow, Russia.
7 - Sechenov University. Moscow, Russia.
8 - S.S. Yudin City Clinical Hospital of the Moscow City Health Department. Moscow, Russia.
|
P |
urpose. To train, validate and conduct interim testing of a deep learning-based neural network for the automated reconstruction of the prostate using multiparametric magnetic resonance imaging (mpMRI).
Materials and Methods. From January 10, 2023 to October 1, 2023 MRI data were retrospectively and prospectively collected from patients who underwent treatment for prostate cancer, specifically radical prostatectomy. The data annotation was performed manually by three experts with over 10 years of experience in independently conducting and interpreting mpMRI of the pelvis according to the PIRADS v2/2.1 protocol, working independently from each other. During the study, four architectures were trained and tested: nnU-Net, SegResNet, TransBTS, and Swin UNETR. The Dice metric was used to evaluate the accuracy of prostate gland reconstruction during training, validation and testing.
Results. Data meeting the inclusion criteria were collected from five institutions, totally 739 studies. After preliminary testing, the Swin UNETR architecture was chosen as the primary model. The final Dice metric accuracy for prostate reconstruction during training, validation and testing was 0.94, 0.93, and 0.91, respectively.
Discussion. Morphometric measurements of the prostate play a crucial role in differential diagnosis of various pathologies and determining the clinical significance of prostate cancer (PCa). Despite having a standardized protocol for conducting and interpreting studies, there remains significant subjective influence by specialists when interpreting MRI data, including determining prostate volume. This issue is further complicated by individual factors that distort the normal geometry of the organ, minimizing the adequacy of assessments using an ellipsoid-based formula. Currently, artificial intelligence (AI) is one of the most promising tools for automating routine measurements of various body structures, including the prostate. This is supported by numerous international studies demonstrating the feasibility of this concept. Unfortunately, research on AI application for automated prostate contouring using mpMRI data is lacking in Russia, despite global interest in this technology. This study provides a description of the main stages of implementation and preliminary testing results of a domestic second opinion system based on deep learning for PCa diagnosis using mpMRI data, specifically an automation module for prostate contouring. Considering the strategy for data collection and annotation, as well as testing multiple neural network architectures to identify the most suitable solution for the tasks at hand, the results are promising. This justifies further research to determine the true advantages of the proposed second opinion system in clinical practice.
Conclusion. The presented results of developing a domestic AI-based second opinion system for prostate reconstruction using MRI data are promising given its development protocol. Despite sufficient accuracy, further multicenter studies are needed to determine the clinical applicability and effectiveness of this solution for integration into routine specialist practice.
Keywords: prostate gland, magnetic resonance imaging, contouring, volume, artificial intelligence.
Corresponding author: Talyshinskii A.E., e-mail: Этот e-mail адрес защищен от спам-ботов, для его просмотра у Вас должен быть включен Javascript
For citation: Talishinskii A.E., Kamyshanskaya I.G., Kryuchkova O.V., Bakhtiozin R.F., Pashkovskaya A.A. Application of deep learning for automated prostate contouring on magnetic resonance imaging data: preliminary results. REJR 2024; 14(4):179-188. DOI: 10.21569/2222-7415-2024-14-4-179-188.
Received: 01.10.24 Accepted: 20.11.24