BACKGROUND

Trophoblast cell surface antigen 2 (TROP2) is a promising target for various cancers, including breast cancer. The development of noninvasive techniques for assessing TROP2 expression in tumors holds considerable importance. This study aims to explore the efficacy of machine learning models based on multi-b-value diffusion-weighted imaging (DWI) using the stretched-exponential model (SEM) for predicting TROP2 expression in breast cancer in nude mouse models.

MATERIALS AND METHODS

Thirty-two nude mouse breast cancer models were subjected to 1.5T magnetic resonance imaging (MRI). Using the freely available software package FireVoxe, we extracted the distribution diffusion coefficient (DDC) and water molecule diffusion heterogeneity index (α) values from SEM, along with histogram parameters of DDC and α maps. TROP2 expression was identified by immunohistochemical staining, with integrated optical density (IOD) quantifying the expression levels. Mice were categorized into high and low TROP2 expression groups based on the median IOD. Key imaging parameters were selected to establish three machine learning models: extreme gradient boosting (XGBoost) classifier, logistic regression, and adaptive boosting (AdaBoost) classifier. We compared the models using the area under the curve (AUC) of the receiver operating characteristic (ROC) on a validation set to determine the superior model. The dataset was split into a training set (28 cases) and a test set (4 cases). The selected model was trained to optimize its performance. We evaluated the models' predictive accuracy in estimating TROP2 expression using AUC, calibration curve, and decision curve analysis (DCA).

RESULTS

Thirty-eight imaging parameters, including DDC, α value, and 36 histogram parameters, were extracted per sample. Using these, we identified eight key imaging parameters for constructing the machine learning models. The validation set AUC values for the XGBoost, logistic regression, and AdaBoost models were 0.828, 0.639, and 0.728, respectively, with XGBoost demonstrating superior prediction performance. In the training set, XGBoost achieved an AUC of 1, sensitivity of 0.911, specificity of 1, and accuracy of 0.954; each of these values was 1 in the test set. Cross-validation yielded an AUC of 0.689, sensitivity of 0.567, specificity of 0.567, and accuracy of 0.580. The calibration curve's Brier score was 0.044, indicating proximity to the ideal curve. DCA indicated favorable net benefits within a risk threshold range of 20-90%.

CONCLUSIONS

Machine learning models based on SEM show promise for predicting TROP2 expression in breast cancer in nude mouse models. Among the models, XGBoost demonstrated outstanding performance, suggesting its potential for clinical applications.