Research Information System
Journal of the Faculty of Engineering and Architecture of Gazi University, vol.39, no.3, pp.1401-1416, 2024 (SCI-Expanded)
The Mask R-CNN uses DICOM data as input for segmentation of prostate zones. Segmentation results are saved in JSON file. The results are transferred to PACS servers using the Flask library based on the TCP/IP protocol. The outputs of segmentation are visualized. The block diagram is shown in Figure A. (Figure Presented) Purpose: The main purpose of this study is to perform fully automatic segmentation of prostate zones. While providing this, it is aimed to save the time while reducing the workload on health employees. Theory and Methods: The need for decision support systems in the field of health is increasing every year. One of the most important methods used in the building of this decision support system is artificial intelligence (AI). The use of computer vision, which is a sub-research area of artificial intelligence, in health is vital. Computer vision in the health area consists of four main tasks, image classification, localization, object detection, and image segmentation. In this context, researches related to segmentation were carried out in this study. The Mask R-CNN with different backbone models was used for segmentation of the prostate zones while examining prostate magnetic resonance imaging (MRI) images of 15 patients. Results: The X101-FPN (includes ResNext-101 and Feature Pyramid Network) with a 3x schedule and the R50-DC5 (makes fine tuning starting from the fifth convolution layer for ResNet-50, and includes dilated convolution layer) with a 3x schedule achieved equal performance with a mAP50 value of 96.040, while the training time of the R50-DC5 with a 3x schedule is 16.31 minutes, and the training time of the X101-FPN with a 3x schedule is 32.32 minutes. Therefore, the usability of the R50-DC5 with a schedule of 3x is higher because it gives faster segmentation results in the testing phase. Conclusion: AI model was integrated into the PACS system as ready-to-use in hospitals. In accordance with PI-RADS v2, the segmentation of prostate zones, which is important in the diagnosis of prostate cancer, was automated.