Researchers developed a new model, the TN-USMA model, to aid the distinction between high-risk and low-risk gastrointestinal stromal tumors (GISTs). “It is accurate, robust, generalizable, and efficient for potential clinical applications,” Liu et al stated.
To evaluate the method, researchers included 914 patients with GIST from 18 hospitals in China. In total, they analyzed 1824 EUS images. The diagnostic performance of the model was compared with a radiomics-based method and 2 state-of-the-art deep learning methods.
“Accurate quantification of gastrointestinal stromal tumors’ (GISTs) risk stratification on multicenter endoscopic ultrasound (EUS) images plays a pivotal role in aiding the surgical decision-making process,” the researchers wrote.
In terms of performance, the TN-USMA model achieved an overall accuracy of 0.834 (95% CI, 0.772-0.885), a sensitivity of 0.844 (95% CI, 0.672-0.947), and a specificity of 0.832 (95% CI, 0.762-0.888). The area under the receiver operator curve (AUC) was 0.881 (95% CI, 0.825-0.924).
Read more about GIST diagnosis
The model developers found that the performance was stable under different variations of multicenter dataset partitions. In addition, the AUC performed better than 2 other deep learning approaches (P <.05).
As explained by the study authors, the TN-USMA model is based on a triple normalization-based deep learning framework with ultrasound-specific pretraining and meta attention. The triple normalization ensures the normalization of intensity, size, and spatial resolution of endoscopic ultrasound (EUS) images. This generates same-size regions of interest and same-resolution tumor masks. To deal with data scarcity, Liu et al used the transfer learning strategy.
The next step involves feeding the same-size regions of interest into a deep architecture with ultrasound-specific pretrained weights. Measurements of tumor size are calculated using the same-resolution masks. Ultimately, size features and 2 demographic features are added to the model prior to the final classification layer using a meta attention mechanism. This step increases feature representations.
Liu C, Qiao M, Jiang F, Guo Y, Jin Z, Wang Y. TN‐USMA Net: triple normalization‐based gastrointestinal stromal tumors classification on multicenter EUS images with ultrasound‐specific pretraining and meta attention. Med Phys. Published online August 19, 2021. doi:10.1002/mp.15172