Fully automated pelvic bone segmentation in multiparameteric MRI using a 3D convolutional neural network.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate segmentation of pelvic bones is crucial for detecting pelvic bone metastases.
• Current methods can be time-consuming and require expert input.
• Deep learning offers a potential solution for automated segmentation.

Purpose of the Study:

• To develop and evaluate a deep learning-based approach for automated segmentation of normal pelvic bony structures.
• To utilize a 3D convolutional neural network (CNN) for this task.
• To assess the performance of the automated segmentation on multiparametric MRI (mpMRI) data.

Main Methods:

• A retrospective study of 264 pelvic mpMRI datasets (2018-2019) was conducted.
• Manual annotations of pelvic bony structures served as reference standards.
• A 3D U-Net CNN was trained and validated on diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) images.
• An external validation set of 60 mpMRI scans (2020) was used.

Main Results:

• The CNN achieved high Dice similarity coefficient (DSC) scores: 0.80 (DWI) and 0.85 (ADC) on the testing set, and 0.79 (DWI) and 0.84 (ADC) on the external validation set.
• Pelvic bone volumes showed strong correlation (R² 0.84-0.97) and agreement (mean bias 2.6-4.5 cm³) between manual and CNN segmentations.
• Qualitative evaluation using a SCORE system demonstrated high performance and inter-reader agreement (ICC = 0.904).

Conclusions:

• A deep learning-based method effectively automates pelvic bone segmentation on DWI and ADC images.
• The developed 3D CNN demonstrates suitable quantitative and qualitative performance.
• This automated approach can support accurate detection and localization of pelvic bone metastases.