GST is the most common gastrointestinal stromal tumor (GIST), often occurring in middle-aged and elderly people, and with similar incidence rates in men and women [9]. The clinical symptoms are not specific, and, because only a small amount of submucosal tissue can be obtained by preoperative biopsy, it is difficult to accurately judge tumor heterogeneity [20]. In addition, preoperative biopsy is a common cause of tumor rupture and bleeding, leading to an increased risk of tumor dissemination. CT is an effective tool to support the diagnosis and differential diagnosis of gastrointestinal diseases [21]. GST and GS have similar clinical and imaging manifestations, whlie frequently difficult to discriminate. Therefore, non-surgical approaches to distinguish these two tumors are needed to ensure proper clinical management. In this study, a large number of clinical and CT features were inductively analyzed, the best features to identify the two tumor types were screened, and the three top-scoring features were input into five ML algorithms to construct prediction models to find the best model to distinguish small GST from GS.

Regardless of whether the dependent variable is continuous or categorical, LASSO regression can be applied by constructing a penalty function (λ) to eliminate low-correlation features and retain the optimal high-correlation features. In this study, five optimal features were screened out, including HDv, lobulation, peripheral lymph nodes, DEd, and tumor growth site.

HD is the standard deviation of CT in the tumor, and it is an indirect reflection of intratumoral heterogeneity. Tumors with different pathological bases have different degrees of heterogeneity ≤ 5 cm GSTs are low-grade malignant tumors or tumors with malignant potential, whereas GSs are almost always benign, and we found that these two GSMT subtypes are characterized by different HDs. However, only HDv qualified for inclusion in the prediction models, which may be owing to the fact that the amount of detectable heterogeneity between the two tumor types varies according to the post-enhancement phase, and that the heterogeneity between tumors is most prominent in the venous phase.

Malignant tumors grow faster than benign tumors, and the difference in proliferation rate of tumor cells results in irregular lobulated changes in the tumor body. Moreover, the higher the degree of malignancy, the greater the probability of lobulation [22]. GSs contain varying numbers of inflammatory cells, so reactive lymph nodes of different sizes often appear around the lesion, whereas ≤ 5 cm GSTs do not contain inflammatory cells, so this phenomenon is relatively rare. On enhanced scans, both tumors showed progressive enhancement, but the degree of progressive enhancement of ≤ 5 cm GSTs was lower than that of ≤ 5 cm GSs, and the peak was more anterior [23]. Thus the DE of the two tumor types after enhancement is different, and the difference becomes more obvious over time, which probably contributes to DEd being screened as an optimal feature. ≤ 5 cm GST and GS also have different predilection sites, with the former being likely to occur in the gastric body and gastric fundus [24], and the latter most commonly occurring in the gastric body followed by the gastric antrum and gastric fundus [25]. Univariate and multivariate logistic regression analysis of the five potential risk factors screened out using LASSO showed that HDv, lobulation, and tumor growth site were independent risk factors. This indicates that peripheral lymph nodes and DEd have some value in distinguishing small GSTs from GSs, but the value is limited.

The prediction models constructed by all five ML algorithms for the differential diagnosis of ≤ 5 cm GST and GS showed high efficiency. There have been other reports of using CT imaging data for the differential diagnosis of GISTs. Sun Jun [26] used the CT whole tumor histogram to identify 6 highly correlated histogram parameters; ROC curve was used to analyze the diagnostic efficiency of statistically significant parameters, and the highest AUC was 0.78. Wang Jian [27] used CT image features to identify ≤ 5 cm GST and GS, and found highly correlated features, ROC analysis resulted in a maximum AUC of only 0.674. The diagnostic efficiency and sample size of these previous studies were lower than this study. Our study incorporates new variables and uses a variety of ML algorithms to build an effective prediction model with improved performance on both the training set and the test set, which indicates the model is generalizable to other clinical samples. Wang [28] used CT images and ML to identify GSTs and GSs and found that the model constructed by logistic regression in the test set had the highest diagnostic efficiency, with an AUC of 0.967. Its diagnostic performance is better than that of ET algorithm in the test set (AUC = 0.912), and lower than that of ET algorithm in the training set (AUC = 0.988). However, it is not completely consistent with the conclusions of this study. In this study, a noninvasive differential diagnosis was performed to distinguish between the GST subgroup (≤ 5 cm) and GS. However, it is important to note that the ET algorithm has inherent limitations. Specifically, when the number of decision trees is substantial, the training time for the model becomes significantly prolonged. Consequently, in practical applications where real-time demands are paramount, the ET algorithm may not be the optimal choice.

This study was subject to several limitations. Firstly, the cases were exclusively obtained from a single hospital, resulting in an inadequate sample size. It is recommended to incorporate multicenter data for future investigations [29], which testing the generalization ability of the model, and gradually apply it in clinical practice [30]. Secondly, this study employed a retrospective analysis, which inherently introduces selection bias. Thirdly, all the included cases of GS were benign, thereby lacking the necessary imaging characteristics of rare malignant GSs. Consequently, the prediction models were unable to evaluate the features of malignant GSs. Lastly, the CT scans were conducted using the empirical method rather than the threshold method, potentially leading to periodic inconsistencies.