A study on the diagnostic performance of a radiomics model based on breast MRI for small breast cancer / 中华放射学杂志

ABSTRACT

Objective:To evaluate the diagnostic performance of a radiomics model based on dynamic contrast-enhanced MRI (DCE-MRI) and diffusion weighted imaging (DWI) in small breast cancer (≤ 20 mm in greatest dimension), and to compare the results with those of an experienced radiologist’s interpretation.Methods:A total of 205 small breast lesions in 192 consecutive female patients from June 2016 to January 2018 at Renji Hospital, School of Medicine, Shanghai Jiaotong University, were retrospectively enrolled in the study. All lesions (≤ 20 mm in greatest dimension) were confirmed by surgical pathological results. The lesions were divided into a training set (116 lesions) and an independent test set (89 lesions). Based on preoperative breast DCE-MRI and DWI data, a radiomics model was built using gradient boosting decision tree (GBDT). The GBDT model was applied to the test set for differentiation between malignant and benign small breast lesions. Cases of the test set were also evaluated by an experienced radiologist for benign and malignant diseases differentiation. ROC curve was used to assess the diagnostic performance for the GBDT model and the radiologist evaluation, respectively. Differences in the area under the ROC curve (AUC) were analyzed by the DeLong test. Differences in sensitivity, specificity and accuracy were evaluated by the McNemar test. Kappa values were used to assess the agreement between different evaluation methods.Results:The AUC of the GBDT model (0.950) showed no significant difference from that of the radiologist’s evaluation based on DCE-MRI combing DWI data (0.935) ( Z=0.499, P=0.618). However, it showed the AUC of GBDT model was significantly higher than that of evaluation based on DCE-MRI (0.874) or DWI (0.832) alone ( Z=2.024， P=0.043； Z=2.772， P=0.006). The sensitivity, specificity and accuracy of the best cutoff point of GBDT model were 90.0%, 89.8% and 89.9% respectively. The sensitivity, specificity and accuracy of evaluation based on DCE-MRI combined with DWI were 97.5%, 79.6% and 87.6% respectively. There was no significant difference in diagnostic performance between the two methods (χ 2=0.800,2.286 and 0.083, P>0.05). Conclusions:A radiomics model based on DCE-MRI and DWI images provided good diagnostic performance in small breast cancer. The results of radiomics were favorably comparable with those of experienced radiologist evaluation based on the combination of DCE-MRI and DWI data.