A automatic segmentation model of bone lesion in bone SPECT/CT based on deep learning
10.3760/cma.j.cn321828-20241106-00384
- VernacularTitle:基于深度学习的SPECT/CT骨显像病灶区域自动分割模型开发
- Author:
Xueting WANG
1
;
Weiming XIE
;
Yujia MIAO
;
Zhaomin YAO
;
Yingxin DAI
;
Fengmin LIU
;
Guoxiu LU
;
Guoxu ZHANG
;
Zhiguo WANG
Author Information
1. 北部战区总医院核医学科，沈阳 110016
- Publication Type:Journal Article
- Keywords:
Neoplasm metastasis;
Skeleton;
Deep learning;
Image processing, computer-assisted;
Positron-emission tomography;
Tomography, X-ray computed;
Technetium Tc 9
- From:
Chinese Journal of Nuclear Medicine and Molecular Imaging
2025;45(11):666-671
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop a deep learning-based segmentation model MT-UNet to automatically segment bone metastases and benign bone lesions in bone scintigraphy with SPECT/CT.Methods:A total of 93 patients (48 males and 45 females, age 28-84 years) who underwent bone SPECT/CT in the Department of Nuclear Medicine, General Hospital of Northern Theater Command from June 2023 to December 2023 were enrolled retrospectively in this study, with a total of 184 bone lesions (94 benign lesions and 90 metastatic tumors). The MT-UNet was employed to segment bone lesions in SPECT, CT and SPECT/CT images respectively. Comparative analysis with 8 segmentation models was performed. The training set and validation set were divided by using 5-fold cross-validation and transfer learning was introduced to further enhance the robustness of the model. An additional cohort of 22 patients (15 males and 7 females, age 37-87 years) who received bone SPECT/CT in the Department of Nuclear Medicine, General Hospital of Northern Theater Command from April 2023 to May 2023 were included, comprising 40 bone lesions (22 benign lesions and 18 metastatic tumors) as the test set of MT-UNet. Segmentation performance of different models was assessed using accuracy, sensitivity, specificity, AUC, intersection over union and Dice similarity coefficient (DSC). Delong test was used to compare the segmentation efficacy among different models in the test set.Results:In the validation set, MT-UNet demonstrated DSC of 0.940, 0.962, and 0.963 for SPECT, CT, and SPECT/CT bone lesion segmentation, respectively, which were outperformed other models. Following transfer learning implementation, the SPECT/CT model′s DSC was improved to 0.984. In the test set, MT-UNet maintained comparable segmentation performance to the validation set, with significant AUC differences among the three models ( Z values: from -15.42 to -9.27, all P<0.01). Compared with conventional image interpretation, MT-UNet-based segmentation reduced physician interpretation time from 164min to 102min. Conclusion:MT-UNet has shown good performance in automatic segmentation of bone metastases and benign bone lesions, and is expected to become an important part of SPECT/CT image intelligent diagnosis system for bone metastases.
