MATU-Net: A Deep Learning-Based Pipeline for Auto Segmentation of Pelvic Lymph Nodes During Brachytherapy of Locally Advanced Cervical Cancer
E Showkatian1*, Z Siavashpour2, S Sadeghi3, (1) Department of Medical Physics, School of Medicine, Iran University of Medical Sciences, Tehran, Iran, (2) Department of Radiotherapy Oncology, Shohada-e Tajrish Medical Center, School of Medicine, Shahid Beheshti University of Medical Science, Tehran, Iran, (3) Department of Nuclear Physics, Faculty of Sciences, University of Mazandaran, Babolsar, Iran
Presentations
PO-GePV-I-53 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: Computed tomography (CT)-based manual delineation of pelvic lymph nodes (LNs) during image-guided brachytherapy (IGBT) of locally advanced cervical cancer is a challenging task. This study aimed to develop a two-step hierarchical convolutional neural network (CNN) approach for the automatic segmentation of LNs from CT images of patients' pelvic during IGBT of locally advanced cervical cancer, focusing on the geometric evaluation.
Methods: A hierarchical coarse-to-fine segmentation approach was utilized to delineate pelvic nodal chains including right/left internal iliac, external iliac, obturator, and pre-sacral. In the coarse segmentation stage, an organ-specific region of interest (ROI) localization map was generated using 3D U-net. In the fine segmentation stage, detailed and accurate contouring of the LNs were produced using Mask R-CNN architecture. The dataset consisting of CT image series from 85 locally advanced cervical cancer patients was randomly divided into the training set (n = 65), the validation set (n = 10), and the test set (n = 10). The model’s performance was assessed by computing the Dice similarity coefficient (DSC), and Hausdorff distance (HD), and 95% HD between the automatically and manually (i.e. ground truth) generated contours on 10 CT scans
Results: The mean DSC, HD, and 95% HD of our hierarchical CNN-based tool were 91.9%/0.82 mm/5.52 mm for the right internal iliac, 88.7%/0.83 mm/5.65 mm for the left internal iliac, 91.9%/0.85 mm/5.97 mm for the right external iliac, 91.5%/0.79 mm/5.78 mm for the left external iliac, 92.0%/0.98 mm/5.75 mm for the right obturator, 90.7%/0.76 mm/4.80 mm for the left obturator, and 82.2%/1.06 mm/4.98 mm for the pre-sacral LNs.
Conclusion: The experimental results show the efficacy of the proposed deep learning-based hierarchical auto-segmentation model for automatic pelvic LN delineation on CT images, which can potentially reduce intra-observer organ delineation uncertainties during IGBT of locally advanced cervical cancer.
Keywords
Taxonomy
Not Applicable / None Entered.
Contact Email