ESTRO 2024 - Abstract Book

S2989

Physics - Autosegmentation

ESTRO 2024

88

Digital Poster

Automatic lymph node clinical target volumes delineation for NPC radiotherapy: A multi-center study

Wenjun Liao 1 , Xiangde Luo 2 , Shichuan Zhang 1

1 Sichuan Cancer Hospital, Radiotherapy, Chengdu, China. 2 University of Electronic Science and Technology of China, Mechanical and Electrical Engineering, Chengdu, China

Purpose/Objective:

The problem of obtaining accurate and individual lymph node (LN) clinical target volumes (CTVs) for nasopharyngeal carcinoma (NPC) radiotherapy (RT) with deep learning (DL) remains unsolved. Herein, we developed a DL model specifically tailored for LN CTVs delineation in NPC.

Material/Methods:

We retrospectively collected 262 patients with 440 planning computed tomography (CT) images from four tertiary hospitals. A total of 13 LN CTVs were designed to adapt to all clinical scenario in NPC RT according to LN involvement. This included left (L)_Ib, L_II+III+Va, L_IV+Vb+Vc, right (R)_Ib, R_II+III+Va, R_IV+Vb+Vc, L_Ib-V, L_II-V, R_Ib-V, R_II-V, bilateral (B)_II+III+Va, B_Ib-V, and B_II-V. Our model was trained on 240 CT images and the performance of our model was validated on 152 CT images across three external testing cohorts. To further assess the clinical accuracy of our model, a separate evaluation was conducted on 48 paired treatment-related CT images from another two external cohorts. Besides, the accuracy of our model was compared with four previous algorithms. The Dice similarity coefficient (DSC) and 95th-percentile Hausdorff distance (HD95) were calculated to evaluate the accuracy for each LN CTV. Clinical accuracy evaluation was classified into three levels: clinical acceptable, acceptable with minor edits, and unacceptable.

Results:

In the internal testing cohort, 84.6% (n = 11) of LN CTVs achieved a mean DSC ranging from 84.49% to 88.46% and all LN CTVs exhibited a mean HD95 below 5 mm, ranging from 3.56 mm to 4.80 mm. Consistent accuracy was obtained in two external testing cohorts, with 11 out of the 13 LN CTVs achieving a mean DSC higher than 83.20% and all LN CTVs having a mHD95 below 5 mm. Moreover, no significant difference was observed in the segmentation of LN CTVs between post-treatment and treatment-naïve CT images, as indicated by both DSC and HD95 measurements. When compared our method with the four algorithms among the five testing cohorts, the average DSC of our method was significantly higher and the average HD95 was significantly lower.When experts evaluated auto-segmentations from the randomly selected 25 patients, 96.3% of auto-segmentations were either scored as being clinically acceptable or requiring minor edits.

Conclusion: