T Peng, K Wang, M Dohopolski, H Shao, Y Zhang, J Wang*, UT Southwestern Medical Center, Dallas, TX

• 
  J Wang

TU-I345-IePD-F5-4 (Tuesday, 7/12/2022) 3:45 PM - 4:15 PM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 5

Purpose: Accurate lymph node (LN) segmentation is pivotal for quantitative assessment of disease progression and potential therapeutic target identification in head and neck cancer. However, the accuracy of LN segmentation is limited due to low contrast between LNs and surrounding tissues, and a large variation of LNs size. The purpose of this study is to develop an automatic LN segmentation method based on deep parallel squeeze & excitation and attention 3D Unet network (SEA-UNet).

Methods: Contrast-enhanced CT images of total 603 LNs from 103 patients with oropharyngeal squamous cell carcinoma were utilized in this study. In this work, we incorporated squeeze & excitation (SE) and attention gate (AG) modules into a popular 3D U-Net architecture to obtain accurate LN segmentation through boosting useful features, while suppressing weak ones. As both SE and AG modules have a good ability to boost useful features for the final LN segmentation, we explored four variants using both SE and AG modules, where two variants are in serial and another two are in parallel architectures. Each model was trained using randomly selected 452 LNs and tested using the remaining 151 LNs. Dice loss was used as the loss function. The accuracy of the segmentation results was quantitatively evaluated using Dice similarity coefficient (DSC) and Hausdorff distance (HD).

Results: The SEA-UNet outperformed several other state-of-the-art approaches for LN segmentation. The parallel architectures (between SE and AG) achieved higher DSC (mean 0.833) and lower HD (mean 3.71 mm) than those of serial architectures with mean DSC of 0.795 and HD of 4.04 mm.

Conclusion: We developed a 3D SEA-UNet for cervical lymph node segmentation in CT images. The parallel structure between SE and AG modules achieved the best performance among all evaluated models.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email