MRI-based radiomic and machine learning for prediction of lymphovascular invasion status in breast cancer

Abstract

Objective: Lymphovascular invasion (LVI) is critical for the effective treatment and prognosis of breast cancer (BC). This study aimed to investigate the value of eight machine learning models based on MRI radiomic features for the preoperative prediction of LVI status in BC.

Methods: A total of 454 patients with BC with known LVI status who underwent breast MRI were enrolled and randomly assigned to the training and validation sets at a ratio of 7:3. Radiomic features were extracted from T2WI and dynamic contrast-enhanced (DCE) of MRI sequences, the optimal feature filter and LASSO algorithm were used to obtain the optimal features, and eight machine learning algorithms, including LASSO, logistic regression, random forest, k-nearest neighbor (KNN), support vector machine, gradient boosting decision tree, extreme gradient boosting, and light gradient boosting machine, were used to construct models for predicating LVI status in BC. The area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity were used to evaluate the performance of the models.

Results: Eighteen radiomic features were retained to construct the radiomic signature. Among the eight machine learning algorithms, the KNN model demonstrated superior performance to the other models in assessing the LVI status of patients with BC, with an accuracy of 0.696 and 0.642 in training and validation sets, respectively.

Conclusion: The eight machine learning models based on MRI radiomics serve as reliable indicators for identifying LVI status, and the KNN model demonstrated superior performance.This model offers substantial clinical utility, facilitating timely intervention in invasive BC and ultimately aiming to enhance patient survival rates.

Keywords: Breast cancer; Lymphovascular invasion; Machine learning; Radiomics.

Fig. 1

Patient selection flowchart

Fig. 2

Workflow of the radiomics analysis

Fig. 3

Radiomics feature selection using the LASSO algorithm. The cross-validation curves of LASSO regression analysis (A) and the LASSO coefficient path plots (B) reveal the radiomics feature selection procedure in the DCE sequence; (C). and (D) represent the radiomics feature selection procedure in the T2WI sequence

Fig. 4

The ROC curves demonstrate the discriminatory performance of the LASSO, LR, RF, KNN, SVM, GBDT, XGBoost, and LightGBM models to predict LVI in BC patients. (A) in the training set. (B) in the validation set