Researchers from China developed a multiscale image normalization method that could improve the classification of gastrointestinal stromal tumors (GISTs) on endoscopic ultrasound (EUS) images.
“The proposed framework can automatically extract discriminative features from multi-center data without requiring any expert knowledge for defining features,” the authors wrote in the study published in Technology and Health Care. Hence, this model might help solve the problem of multicenter data bias.
By using the new proposed model, the authors could distinguish 32 cases of high-risk tumors from 149 cases of low-risk tumors with an accuracy of 0.796, a sensitivity of 0.813, and a specificity of 0.792. The area under the receiver operator curve was 0.844, which was superior to the clinical diagnosis performance.
Read more about GIST testing
As explained by the authors, the framework explored in this study consisted of 3 parts:
- A multiscale image normalization method that prepared EUS images for subsequent feature extraction by uniformizing their size and resolution
- A multiway feature extraction and fusion module that allowed to obtain distinctive tumor features. For this, the authors built a convolutional neural network model. This step is essential to classify GISTs according to their risk levels since they are visually similar among different risk groups
- An XGBoost classifier to predict the tumor’s risk level
Therefore, this framework assured that tumors’ textural and morphological features were extracted from individual images of the same size and that the size information was calculated from images with the same resolution.
Several studies suggest that deep convolutional neural networks could help solve major image classification problems, but studies on EUS images of GISTs are scarce. Among the current challenges, the authors of this study highlighted data distribution bias of multicenter images, interclass similarity, intraclass variation, and insufficiency of training data.
Liu C, Guo Y, Jiang F, et al. Gastrointestinal stromal tumors diagnosis on multi-center endoscopic ultrasound images using multi-scale image normalization and transfer learning. Technol Heal Care. Published online January 28, 2022. doi:10.3233/THC-228005