Researchers from Korea developed a new system, based on endoscopic ultrasound images and called the EUS-CNN, that can help diagnose gastrointestinal stromal tumors (GISTs).
In the study published in the Journal of Gastroenterology and Endoscopy, the authors stated, “This EUS-CNN can be helpful not only for less-experienced endoscopists but also for experienced ones.”
It can be challenging to differentiate gastric subepithelial tumors such as GISTs from leiomyomas — noncancerous growths of the uterus that often appear during childbearing years — based on endoscopic ultrasound images.
Here, the researchers aimed to develop a reliable object detection model that could differentiate the 2 types of growths.
Read more about GIST diagnosis
To do this, they used 376 images from 114 patients with histologically confirmed gastric GIST or leiomyoma to train their system. They then calculated the accuracy, sensitivity, specificity, and usefulness of their system using a validation set of 170 images from 54 patients. Lastly, they asked 4 endoscopic ultrasound imaging experts and 15 trainees to judge the same validation dataset.
They found that the EUS-CNN had a sensitivity of 95.6% in the per-image analysis. The specificity of the system was 82.1%, while its accuracy was 91.2%. The area under the receiver operating characteristic curve (AUC), a measure of how well a parameter distinguishes between 2 diagnostic groups, was 0.9234. This means that there is a 92.34% chance that the model will be able to distinguish between the 2 diagnoses.
In the per-patient analysis, the sensitivity of the EUS-CNN was 100%. Its specificity was 85.7%, its accuracy was 96.3%, and the AUC was 0.9929.
Importantly, the system “outperformed human assessment in terms of accuracy, sensitivity, and negative predictive value.”
The researchers concluded that their system has a high diagnostic ability to predict a diagnosis of GIST but stated that prospective studies are needed alongside validation in a large multicenter trial.
Oh CK, Kim T, Cho YK, et al. Convolutional neural network-based object detection model to identify gastrointestinal stromal tumors in endoscopic ultrasound images. J Gastroenterol Hepatol. Published online August 8, 2021. doi:10.1111/jgh.15653