Preoperative Prediction of Ki-67 Status in Breast Cancer with Multiparametric MRI Using Transfer Learning
Weixiao Liu1, Yulin Cheng2, Zaiyi Liu3
1Department of Radiology, Guangdong Provincial People's Hospital, Guangdong Academy of Medical Sciences, No.106, Zhongshan 2nd road, Guangzhou 510080 Guangdong, PR China; Graduate College, Shantou University Medical College, Shantou, Guangdong, PR China.
Rationale And Objectives:
Ki-67 is one of the most important biomarkers of breast cancer traditionally measured invasively via immunohistochemistry. In this study, deep learning based radiomics models were established for preoperative prediction of Ki-67 status using multiparametric magnetic resonance imaging (mp-MRI).
Materials And Methods:
Total of 328 eligible patients were retrospectively reviewed [training dataset (n = 230) and a temporal validation dataset (n = 98)]. Deep learning imaging features were extracted from T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and contrast enhanced T1-weighted imaging (T1+C). Transfer learning techniques constructed four feature sets based on the individual three MR sequences and their combination (i.e., mp-MRI). Multilayer perceptron classifiers were trained for final prediction of Ki-67 status. Mann-Whitney U test compared the predictive performance of individual models.
Results:
The area under curve (AUC) of models based on T2WI,T1+C,DWI and mp-MRI were 0.727, 0.873, 0.674, and 0.888 in the training dataset, respectively, and 0.706, 0.829, 0.643, and 0.875 in the validation dataset, respectively. The predictive performance of mp-MRI classification model in the AUC value was significantly better than that of the individual sequence model (all p< 0.01).
Conclusion:
In clinical practice, a noninvasive approach to improve the performance of radiomics in preoperative prediction of Ki-67 status can be provided by extracting breast cancer specific structural and functional features from mp-MRI images obtained from conventional scanning sequences using the advanced deep learning methods. This could further personalize medicine and computer aided diagnosis.
