2997

Fully Automated Pelvic Bones Segmentation in Multiparameter MRI Using a 3D Convolutional Neural Network
xiang liu1, chao han1, and xiaoying wang1
1department of radiology, peking university first hospital, Beijing, China

Synopsis

This retrospective study aims to perform automated pelvic bones segmentation in multiparametric MRI (mpMRI) using 3D convolutional neural network (CNN). 264 pelvic DWI images and corresponding ADC maps obtained from three MRI vendors from 2018 to 2019 were used for the 3D U-Net CNN development. 60 independent mpMRI data from 2020 were used to externally evaluate the segmentation model using quantitative criteria (Dice similarity coefficient) and qualitative assessment (SCORE system). The results demonstrated that the 3D CNN can achieve fully automated pelvic bone segmentation on multi-vendor DWI and ADC images with good quantitative and qualitative performances.

Introduction

Accurate segmentation of pelvic bones is considered as the first step to achieve accurate detection and localization of pelvic bone metastases1. However, the ill-defined boundary among bone structures posed a challenge to the segmentation.

Objective

To present a deep learning-based approach for automated pelvic bones segmentation in multiparametric MRI (mpMRI) using 3D convolutional neural network (CNN).

Methods

This IRB-approved retrospective study included 264 pelvic mpMRI data obtained from three MRI vendors from 2018 to 2019. The manual annotations of pelvic bones (included lumbar vertebra; sacrococcyx; ilium; acetabulum; femoral head; femoral neck; ischium; pubis) on DWI and ADC images were used to create reference standards (Fig. 1). A 3D U-Net CNN was employed for automatic pelvic bones segmentation2. 60 independent mpMRI data from 2020 were additionally used to externally evaluate the model. The segmentation results were assessed quantitatively to the reference standards using the Dice similarity coefficient (DSC). To evaluate whether the predicted results meet the requirements of clinical application. A SCORE system established by us (Fig. 2) was used to qualitatively evaluate the model by two readers, and a rate of up-to-standard >80% was considered to meet the demand of clinical application.

Results

The 3D U-Net achieved high average DSC in both testing and external validation set (testing set: 0.80 [DWI images] and 0.85[ADC images]; external validation set: 0.79 [DWI images] and 0.84 [ADC images]) as shown in Table 1 and Figure 3. In terms of qualitative evaluation, all of the rates of up-to-standard in both testing and external validation set are highly above 80%, and intraclass correlation coefficient (ICC) of the SCORE results between two readers was 0.842 (95% confidence interval: 0.793-0.880) (Table 2).

Discussion

In this research, the quantitative and qualitative evaluations were used to determine the value of clinical application of the 3D U-Net model. The CNN achieved good quantitative segmentation performance on pelvic bones with high average DSC values on DWI images and ADC images in both testing set and external validation set. More importantly, all of the rates of up-to-standard in both DWI and ADC images are highly above 80%, which demonstrated the good acceptability of the 3D U-Net in clinical practice. Besides, the excellent concordance between two readers (ICC = 0.842) further confirmed the feasibility of clinical application of this CNN. Our research may provide essential geometric information for subsequent detection of pelvic bone metastases.

Conclusion

A deep learning-based method can achieve fully automated pelvic bone segmentation on multi-vendor DWI and ADC images with good quantitative and qualitative performances.

Acknowledgements

The authors gratefully acknowledge the technical support of Yaofeng Zhang, Dadou Zhang and Jiahao Huang from Beijing Smart Tree Medical Technology Co. Ltd.

References

1. Lindgren Belal S, Sadik M, Kaboteh R, et al. Deep learning for segmentation of 49 selected bones in CT scans: First step in automated PET/CT-based 3D quantification of skeletal metastases. Eur J Radiol 2019;113:89-95.

2. Çiçek Ö, Abdulkadir A, Lienkamp SS, Brox T, Ronneberger O: 3D U-net: Learning dense volumetric segmentation from sparse annotation. In: Int conf Image Comput Comput Interv. 2016: 424–432.

Figures

Figure 1. Reconstruction of manual segmentations of pelvic bones. A multiclass label was created for the patients: 1, lumbar vertebra; 2, sacrococcyx; 3, ilium; 4, acetabulum; 5, femoral head; 6, femoral neck; 7, ischium; 8, pubis.

Figure 2. The SCORE system and evaluation criteria on DWI and ADC images. Condition A refers to that the location of the predicted CD is consistent with manual CD, and the range of the predicted CD is larger than (A1) or less than (A2) the manual CD, or partially overlaps with the manual CD (A3). CD: connected domain, which is defined as the label with a continuous structure in 3D space. DSC: Dice similarity coefficient.

Figure 3. Examples of the comparison between CNN-predicted and manual segmentation. (a) Section examples of eight bones on DWI image; (b) The corresponding overlapping images between manual segmentation (white background) and CNN-predicted segmentation; (c) Section examples of eight bones on ADC image; (d) The corresponding overlapping images between manual segmentation (white background) and CNN-predicted segmentation .CNN: Convolution neural network.

Table 1 The DSC values of 3D U-Net in pelvic bones segmentation on DWI and ADC images

Table 2 The summary of patient scores on DWI and ADC images

Proc. Intl. Soc. Mag. Reson. Med. 29 (2021)
2997