Head and Neck high-risk lymph nodes detection – a three-dimensional deep learning proposal
PO-1614
Abstract
Head and Neck high-risk lymph nodes detection – a three-dimensional deep learning proposal
Authors: Tuan Yu1, Yihui Lin2,3, YungFa Lu2, YenYu Lin4, ChunRong Huang5, Weisi Yan6
1Taichung Veterans Genearl Hospital, Department of Psychiatry, Taichung, Taiwan; 2Taichung Veterans General Hospital, Department of Radiation Oncology, Taichung, Taiwan; 3National Yang Ming Chiao Tung University, Institute of Computer Science and Engineering, Taichung, Taiwan; 4National Yang Ming Chiao Tung University, Department of Computer Science, Hsinchu, Taiwan; 5National Chung Hsing University, Department of Computer Science and Engineering, Taichung, Taiwan; 6University Of Kentucky, Department of Radiation ONcology, Lexington, USA
Show Affiliations
Hide Affiliations
Purpose or Objective
Nowadays, automated normal organ segmentation by deep learning model is mature, economic efficient and clinically adapted world-wide. When talking about contouring cancerous lesion, it remains state of the art. In our previous works, we successfully recognized metastatic lymph nodes from reactive ones from Head and Neck tomography. To make the workflow more automatically, we experimental several deep learning model to find out best one for high-risk lymph nodes instance detection from pre-treatment tomography.
Material and Methods
We retrospectively collected newly diagnosed head and neck cancer patient. Collect pre-treatment contrast-enhanced tomography. Correlate with pathology report. Label lymph nodes status as (1) metastasis-positive, extranodal-extention-positive (LNM+, ENE+) (2) metastasis-positive, extranodal-extention-negative (LNM+, ENE-) or (3) metastasis-negative (LNM-). If the lymph node did note been dissected, then this node is excluded from our research. If the lymph node cannot be recognized by or neck level, it will be excluded. The recognized lymph nodes then contoured and labels by two Radiation Oncologists. The model performance is evaluated by mean average precision (mAP) under intersection over union (IoU) threshold 0.6.
Results
From 2019-2021, 158 patients with 158 CT scans meet our criteria. The median duration from CT scan to surgery was 10 days (range: 1–28 days). Following pathologic correlation with CT scans, 350 lymph nodes were segmented in total (range: 1–8 per patient): 241 negative nodes, 111 nodes contained tumor cells. In the metastasis-positive nodes, 56 were ENE-positive and 55 were ENE-negative. Diameters of LNM- nodes were 22mm-62mm and LNM+ ones were 28mm-131mm, separately. We set the high-risk threshold as beyond 28mm diameters in any axis. In our dataset, 234 nodes assigned as high-risk and 116 were low-risk. After data augmentation, the dataset were divided into train, validation and test set as 7:1:2 ratio. The mAP of hierarchical LSTM was 0.89.
Conclusion
Hierarchical LSTM based deep learning model served as a useful tool for head and neck lymph node detection. Further external validation is needed before deployed into clinical use.
This work was supported in part by the National Science and Technology Council, Taiwan under Grant of NSTC 111-2634-F-006-012