In this study, we retrospectively collected patients' clinical data and screened for potentially influential independent variables. Among these variables, pronounced redness is one of the characteristics that is often defined subjectively, as it is susceptible to observer bias due to differences in the color and brightness level of the surrounding environment under the endoscope. was. To solve such challenges, in this study he innovatively used CIE to differentiate between chromaticities. To our knowledge, such an approach has not been applied to his EGC depth prediction model.
Regarding univariate analysis of all independent variables collected, it was demonstrated that lesions with greater color difference from surrounding tissue tended to have a higher risk of deeper invasion. Such findings suggest that being able to measure her WLI color metrics in real time on endoscopic images may improve the diagnostic accuracy of EGC depth, especially for less experienced endoscopists. suggests that there is. For other predictive features, Abe et al.2 and Choi et al.3 reported length ≥ 30 as an independent risk factor for deep invasion, which is consistent with our results.Lesions located in the upper third of the stomach are more likely to invade the SM layer.4,17, is thought to be related to the thin structure of the gastric wall in the upper part of the stomach. In addition, upper lesions are more difficult to detect in the early stages due to the limited viewing angle. In Abe et al., marked elevation of the margin was considered to be a promising predictor.2Nagahama et al.13and the study of Yao et al.18. After precisely defining the factors, a single use of the indicator of significant marginal elevation can achieve a sensitivity of 92%, specificity of 97.7%, and accuracy of 96.9%.13. When cancer cells infiltrate the submucosal layer, cancer cell clusters and fibrosis at the infiltrated site cause local stiffness and hypertrophy at the submucosal infiltrated site, and when the gastric wall is completely expanded by endoscopic air supply. , the submucosal infiltration site has not spread, but the surrounding area has spread and shows a marginal protuberance.Yamada et al.19 and Jiang et al.20 Lesions reported as depressed type and mixed histology were more prone to SM invasion and lymph node metastasis, which was consistent with our results.
Based on the variables screened from the univariate analysis, we trained a logistic regression and built a nomogram model as a benchmark. The logistic regression model reached an AUROC of 0.840 on the validation set. To improve prediction accuracy and consider clinical applications, we further investigated machine learning algorithms with better model interpretability. A decision tree model and a random forest model were constructed. Both models are better at dealing with nonlinear relationships compared to traditional approaches. And, to the best of our knowledge, this is the first study to focus on these two types of deep learning models for EGC depth prediction. Regarding the clinical implications of each model, the strong model interpretability of decision tree models enabled the design of simple diagnostic procedures. Meanwhile, the random forest model allowed us to understand the importance of clinical indicators. According to our results, the constructed decision tree model can be used to detect abnormalities in margin prominence, lesions located in the lower third of the stomach, a*color value, b*color value of WLI, and contrast-enhanced CT. indicated that a suitable thickness was selected. In the random forest model, six metrics have the most influence on the prediction results: margin height, WLI a* color value, WLI color difference, WLI b* color value, and location in the middle or lower part of the stomach. The factors screened in all three of his models developed in this paper are generally consistent, with only slight differences in the predictive importance of some variables. Among all three models, Random Forest was dominant with an AUROC of 0.844. The machine learning algorithm also suggested that WLI b* color values and enhanced CT could improve prediction accuracy, which warrants further investigation.
Apart from the aforementioned decision tree models and random forest models, scholars have also studied the application of CNNs in EGC depth prediction. Yun et al.Ten developed a CNN model with an AUROC of 0.851 using 11,539 endoscopic images. Zhu et al.12 and Nagao et al.twenty one reported a CNN model that achieved an AUROC of 0.94 and 0.959, respectively. Goto et al.9 We developed a diagnostic method using an endoscopist and an AI classifier, and achieved an accuracy of 78.0%, which is higher than the AI classifier or endoscopist alone. Although his AUROC and accuracy of CNN appear to be higher than the decision tree and random forest models considered in this study, his CNN model of decision-making progress is more like a black box with low interpretability. Thing. Some scholars say that CNN models exhibit a tendency to overfit. Decision tree models provide a clear decision-making process that is easy to understand clinically, and random forest models visualize the importance of each feature in a predictive model, so both are easy to understand and can be used at different levels of clinical practice. is easier to apply. Current deep learning models for EGC depth prediction mainly use still images rather than videos, which is different from clinical practice. Wu et al.11 attempted to introduce real-time video for gastric cancer lesion detection in a deep learning model and reached a sensitivity of 92.8%. Real-time video may be further applied to machine learning of gastric cancer invasion depth in future studies.
The main innovations of this study are: First, we investigate multiple variables including clinical features, laboratory tests, CT results, endoscopic features, and pathological results. Among them, he innovatively introduced CIE to quantify color. This standardizes color metrics and eliminates subjectivity. This result is of great clinical value, and by installing a plug-in applet within an endoscopic imaging system, the endoscopist can automatically calculate the color difference between selected regions in real time, allowing Helps estimate the depth of infiltration. The patient. Next, apart from the logistic regression model, we further introduce machine learning into the study. The importance of this feature was intuitively demonstrated by systematically screening a wide range of predictor variables using decision trees and random forests. All three models achieved strong predictive results.
This study also has some limitations. First, this was a retrospective, single-center study, which limited the sample size. We attempted to establish cutoff points for continuous values such as WLI a*color, b*color, and WLI color difference, but the exact cutoff points for these variables will be further explored in future multicenter prospective studies. You need to decide. Second, some images may have limited angles, so there may be some discrepancies in reading historical endoscopic images. Although this study retrospectively reads static images, there is still a gap from dynamic video reading in actual clinical settings. Due to limited sample size, some variables have significant missing data, limiting the deployment of more complex machine learning models. Third, due to the limited number of endoscopically resected submucosal carcinomas, our study included specimens from both surgery and ESD. However, there are some differences in the spacing of the resection sections between these two specimen treatments, which may result in an underestimation of the invasion depth and affect the validity of the predictive model. Future studies with larger samples may attempt to include only endoscopically or surgically resected specimens for more precise analysis. In conclusion, models with color metrics using logistic regression and machine learning algorithms may be useful in determining treatment for EGC. Future prospective studies and external validation could be performed at multiple centers to further validate the accuracy of the model.
