Post written by Jie Tian, PhD, from the CAS Key Laboratory of Molecular Imaging, Beijing Key Laboratory of Molecular Imaging, The State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, and the Beijing Advanced Innovation Center for Big Data-Based Precision Medicine, School of Medicine and Engineering, Beihang University, Beijing, China.
We developed a computer-aided early gastric diagnostic model based on 1777 narrow-band imaging with magnifying endoscopy (ME-NBI) images via deep learning method. The performance of the model was validated in the external cohort and compared with 8 endoscopists with varying experience.
The missing rate of EGC during gastroscopy is about 10%. ME-NBI has become the most powerful tool in evaluating EGC, but it requires substantial expertise and experience, and the diagnosis based on it still often varies widely among experts. Therefore, it is appealing to develop an automatic predictive tool to assist in diagnosing EGC with high efficiency.
Our model (EGCM) acquired the area under the receiver operating characteristic curves of 0.808 in the internal cohort and 0.813 in the external cohort, which showed the generalized ability of our model to other centers with different distributions of the data. It had a similar capability to that of senior endoscopists. After referring to the diagnosis results of EGCM, the diagnostic accuracy of the junior endoscopists was significantly improved from 0.728 to 0.747 (P<0.05) and became comparable to the senior endoscopists while the performance of the senior endoscopists was also significantly improved (P<0.05). These results demonstrated the additional value of EGCM in the auxiliary diagnosis of EGC.
Considering that the ultimate application scenario of AI is a real-time diagnosis based on video streams, a study directly training a model using video data from multiple centers is highly desirable in the future.
We constructed an EGCM model for assisting the EGC diagnosis, which exhibited comparable diagnostic capability to that of expert endoscopists and showed potential in improving the diagnostic performance of non-expert endoscopists.
Figure 4. Comparison between endoscopists and the computer-aided early gastric cancer diagnosis model (EGCM). A, The average performance of endoscopists before and after EGCM assistance. B, Subgroup analysis of the performance differences before and after EGCM assistance. ACC, Accuracy; Sn, sensitivity; Sp, specificity; PPV, positive predictive value; NPV, negative predictive value.
The information presented in Endoscopedia reflects the opinions of the authors and does not represent the position of the American Society for Gastrointestinal Endoscopy (ASGE). ASGE expressly disclaims any warranties or guarantees, expressed or implied, and is not liable for damages of any kind in connection with the material, information, or procedures set forth.