2263 - Deep Learning-Based Automatic Reconstruction of Interstitial Needles in Brachytherapy for Cervical Cancer

02:30pm - 04:00pm PT

Hall F

Screen: 29

POSTER

Presenter(s)

Xianliang Wang, PhD - Sichuan Cancer Hospital & Research Institute, Sichuan, Sichuan Province

S. Wen^1,2, T. Liu^1,2, F. Zeng^1,2, and X. Wang³; ¹Applied Nuclear Technology in Geosciences Key Laboratory of Sichuan Province, Chengdu University of Technology, Chengdu, 610059, China, ²Sichuan Cancer Hospital & Institute, Sichuan Cancer Center, School of Medicine, University of Electronic Science and Technology of China, Radiation Oncology Key Laboratory of Sichuan Province, Chengdu, 610041, China, ³Sichuan Cancer Hospital and Research Institute, University of Electronic Science and Technology of China, Chengdu, China

Purpose/Objective(s):

Needle reconstruction is a crucial step in intracavitary-interstitial brachytherapy (IC-ISBT). Currently, needle reconstruction relies on manual processes, which are time-consuming and prone to errors. Automated segmentation and reconstruction methods for interstitial needles can potentially improve efficiency and reduce clinical workload. The objective of this study is to utilize existing radiation therapy plans to achieve deep learning-based reconstruction of interstitial needles in computed tomography (CT) images of cervical cancer brachytherapy.

Materials/Methods:

98 patients with cervical cancer who underwent CT-guided IC-ISBT were included in this retrospective study, comprising 180 treatment fractions. Interstitial needle masks were automatically generated using the dwell positions of the radiation source. For needle segmentation, a new attention-augmented U-Net++ (AA-U-Net++) model was employed. Image processing and clustering algorithms were used to obtain the central positions of the interstitial needles in the 2D mask. The Random Sample Consensus (RANSAC) algorithm was applied for interstitial needles classification, achieving 3D reconstruction. The model's performance was assessed using the Dice similarity coefficient (DSC) and Jaccard index (JI). The geometric deviation of the results was evaluated using tip error and shaft error. The efficiency of the method was assessed by recording the time taken for automatic reconstruction. The results were also compared with those from the standard U-Net model and the U-Net++ model.

Results:

AA-U-Net++ outperformed both U-Net++ and U-Net, achieving a DSC of 94.9 ± 1.9 and a JI of 91.0 ± 2.6. U-Net++ scored lower (DSC: 93.6 ± 2.2; JI: 90.2 ± 2.6), while U-Net showed the least accuracy (DSC: 91.9 ± 3.0; JI: 87.2 ± 4.0). AA-U-Net++ also demonstrated superior precision in needle localization, with mean tip and shaft errors of 0.29 ± 1.60 mm and 0.44 ± 0.52 mm, respectively. These errors were smaller than those of U-Net (tip: 0.72 ± 2.02 mm; shaft: 0.65 ± 0.59 mm) and U-Net++ (tip: 0.38 ± 1.96 mm; shaft: 0.45 ± 0.53 mm). The maximum errors for AA-U-Net++ were 3.28 mm for tip and 2.79 mm for shaft. Notably, U-Net failed to reconstruct two interstitial needles, while U-Net++ showed higher maximum errors (tip: 9.24 mm; shaft: 5.42 mm).

Conclusion:

The proposed method can accurately reconstruct interstitial needles in CT images for IC-ISBT by leveraging the dwell position of radiation sources, enhancing the efficiency of the procedure.