51. An Artificial Intelligence System for Detecting Esophageal Squamous Cell Carcinoma and Precancerous Lesions Using Endoscopic Images and Videos.
- Author
-
Yoonhee Lee, Seunghan Lee, Jiwoon Jeon, Jinbae Park, Cheol Min Shin, Mi Jin Oh, San Gyun Kim, and Soo-jeong Cho
- Subjects
CONVOLUTIONAL neural networks ,RECEIVER operating characteristic curves ,SQUAMOUS cell carcinoma ,ARTIFICIAL intelligence ,PRECANCEROUS conditions - Abstract
Background/Aims Endoscopic screening for early detection of esophageal squamous cell carcinoma (ESCC) is essential due to the poor prognosis of advanced ESCC. This study aimed to develop an artificial intelligence (AI) model based on a convolutional neural network to accurately detect and diagnose ESCC and its premalignant lesions using endoscopic images and videos. Methods A total of 13,230 endoscopic esophageal images, including white light imaging and narrow band imaging, were collected from 1,864 patients at Seoul National University Hospital between January 2003 and November 2022. The dataset included 7,304 images of esophageal dysplasia, superficial ESCC, advanced ESCC, as well as 4,330 images of benign lesions, and 1,596 images of normal esophagus. To develop the AI system, we randomly divided the dataset into training, validation, and test sets in a 7:2:1 ratio. The model was trained with the training set, internally validated with the validation set, and the performance was evaluated using the test set. Results After training, The model achieved accuracies of 90.83% for cancer detection in validation dataset and 90.95% in test dataset. For the validation dataset, sensitivity, specificity, positive predictive value, and negative predictive value for cancer detection of AI system were 92.87%, 87.76%, 91.93%, 89.13% respectively. For the test dataset, sensitivity, specificity, accuracy, positive predictive value, and negative predictive value were 93.81%, 86.80%, 91.18%, 90.61%. The area under receiver operating characteristic curve for AI system were 0.9594 for validation dataset and 0.9640 for test dataset. Conclusion This AI model demonstrated high sensitivity and specificity for detecting and diagnosing ESCC and dysplasia. This system has the potential to assist endoscopists in diagnosing ESCC, which can easily be overlooked, in real-time practice. [ABSTRACT FROM AUTHOR]
- Published
- 2024