362. AUTOMATED DETECTION OF UPPER MEDIASTINAL LYMPH NODES METASTASIS OF ESOPHAGEAL SQUAMOUS CANCER USING BERT ON COMPUTED TOMOGRAPHY
Zhao Ma, Yang Li, Xiangchun Li, Hongjing Jiang- Gastroenterology
- General Medicine
Abstract
Upper mediastinal lymph node (LN) metastasis is the extremely frequent metastasis lymph node site and dissection of this site of lymph node harbors a high risk of recurrent laryngeal nerve injury. Prediction of LN metastasis before dissection is necessary for personalized therapy. We developed a pre-training unsupervised deep learning model based on Bidirectional Encoder Representation from Transformers (BERT) to extract features from pre-operational CT images and detect patients with LN metastasis. A total of 3769 esophageal cancer patients from 2012 to 2020 with complete post-surgical pathology in our institute were included in this study, 823 of which had upper mediastinal LN metastasis. The extracted features of CT images were labeled with the LN metastasis status of the corresponding patients and then divided into training and validation groups (7:3) for training. In order to keep balancing of the positive and negative ratio in the test dataset, 800 patients were randomly selected from the validation group (1:1). The BERT-based pretrained learning model achieved a AUROC of 0.651 on detection of upper mediastinal LN metastasis. The specificity of this model could reach up to 0.732. Our model can be used to improve diagnostic ability of upper mediastinal LN metastasis before surgery.