Background

The status of axillary lymph node (ALN) is closely related to the prognosis of the patients with breast cancer, and determines the mode of surgery and the formulation of adjuvant treatment [1,2]. ALN status is currently determined by invasive SLN biopsy in clinical practice.

Methods

In this retrospective research, we collected 197 patients who underwent breast magnetic resonance dynamic enhancement before operation and had confirmed postoperative pathology for breast cancer from September 2016 to May 2019. According to the postoperative pathology, the patients were divided into ALN-metastasis group(82) and non-ALN-metastasis group(n=115),then the patients were randomly assigned to training group(n=131) and verification group(n=66). A dedicated software (Philips radiomics tool) was used to draw the contour of the tumors in the early phase and late phase enhanced Silhouette images and calculate the features. A total of 3386 radiomic features were extracted from each patient include tumor intensity statistics, size and shape, intensity statistics, and texture feature using pyradiomics [3]. In the following feature dimension reduction analysis, we used Spearman correlation analysis to select the key features. In modeling stage, we investigated 5 classification methods (including Passive Aggressive Classifier, Perceptron, Ridge Classifier, SGD Classifier, Logistic Regression, Linear SVC) for training and prediction. We use 5-fold cross validation results as the performance of a specific machine learning classifier. We use ’accuracy’ as the optimization metric to select the classifier. The model was trained on the training cohort and their performance was evaluated on the cross-validation cohort using the area under ROC curve (AUC).

Results

25 radiomic features were select the key features to predict ALN status, Including 20 features from early phase enhanced images and 5 features from late phase enhanced images. Ridge Classifier was found to produce the most accurate model on training dataset. The prediction model displayed an AUC of 0.82 and 0.79 for predicting ALN status in the training group and validation group, respectively.

Conclusion

Our result shows that some radiomics features have great potential to be an useful index in predicting ALN status, therefore providing helps for the development of clinical treatment decisions for breast cancer.

Discussion

This study tried to predict the ALN metastasis in patients with breast cancer by DCE-MRI radiomic characteristics, and showed good predictive performance in the training group and validation group, respectively (AUC 0.85 ,0.79). MSKCC Nomogram is a model to predict the SLN status which developed by The Memorial Sloan-Kettering Cancer Center. It based on some clinicopathological characteristics, such as tumor size, and pathological type and age, the AUC was 0.754 [4]. Compared with MSKCC Normogram, The predictive efficiency of our radiomic model is better, and as a non-invasive method, which can be carried out before operation, and radiomic will have the potential to provide more auxiliary information for clinical treatment.

Acknowledgements

No acknowledgement found.