3404 - Deep Learning-Based Fully Automated Detection and Segmentation of Lymph Nodes on Computed Tomography for Head and Neck Cancer: A Multi-Center Study

W. Liao¹, X. Luo², and S. Zhang¹; ¹Department of Radiation Oncology, Sichuan Cancer Hospital & Institute, Sichuan Cancer Center; Cancer Hospital affiliate to University of Electronic Science and Technology of China, Chengdu, China, ²School of Mechanical and Electrical Engineering, University of Electronic Science and Technology of China, Chengdu, China

Purpose/Objective(s): Accurate lymph nodes (LNs) assessment is important for head and neck cancer (HNC) staging and treatment decisions. However, it is incredibly time-consuming to identify all LNs on computed tomography (CT). This study aims to develop and validate a deep learning-based, fully-automated LN detection and segmentation (auto-LNDS) model based on CT. Materials/

Methods: In total, 11,013 annotated LNs with a short-axis diameter = 3mm were enrolled on CT images from 663 patients with HNC from four hospitals. The auto-LNDS model first employed nnUNet as a baseline and was pre-trained on a large-scale head and neck organs at risk dataset (more than 5000 patients with more than 45 OARs). Then, the pre-trained model was transferred and finetuned with 4729 annotated LNs from 257 patients of hospital A for automatic LNs detection and segmentation. Afterwards, the model was validated both on internal testing dataset (hospital A) and three external testing datasets (hospitals B, C and D) consisting of 1108 LNs and 5176 LNs, respectively. The performance for LNs detection was evaluated using sensitivity, positive predictive value (PPV), and false positive rate per case (FP/vol), and the segmentation performance was evaluated using the Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95).
Results: For LNs detection, the auto-LNDS achieved sensitivity, PPV, and FP/vol of 54.6%, 69.0%, and 3.4 in internal testing dataset, respectively. And in three external datasets, the auto-LNDS achieved an average sensitivity, PPV, and FP/vol of 50.8%, 59.0%, and 4.5, respectively. The time taken for model detection and segmentation was 1.5 s/case, compared with 560 s/case for the radiologists. For LNs segmentation, the auto-LNDS achieved DSC and HD95 of 0.72 and 3.78 mm in internal testing dataset, respectively. And in three external datasets, the auto-LNDS achieved an average DSC and HD95 of 0.73 and 3.05 mm, respectively. Additionally, there was no significant difference in sensitivity between contrast enhanced CT and unenhanced CT images in internal testing dataset (p = 0.502). And, no significance difference was observed in sensitivity among three groups of repeated CT images obtained during adaptative radiotherapy in hospital D (p = 0.815).
Conclusion: The proposed auto-LNDS model achieves an overall good performance in the automatic detection and segmentation of cervical LNs in multi-center CT images, and holds great potential for facilitating N-staging and radiation treatment planning for HNC.