3T DCE-MRI Radiomics Improves Predictive Models of Complete Response to Neoadjuvant Chemotherapy in Breast Cancer

Abstract

Objectives: To test whether 3T MRI radiomics of breast malignant lesions improves the performance of predictive models of complete response to neoadjuvant chemotherapy when added to other clinical, histological and radiological information.

Methods: Women who consecutively had pre-neoadjuvant chemotherapy (NAC) 3T DCE-MRI between January 2016 and October 2019 were retrospectively included in the study. 18F-FDG PET-CT and histological information obtained through lesion biopsy were also available. All patients underwent surgery and specimens were analyzed. Subjects were divided between complete responders (Pinder class 1i or 1ii) and non-complete responders to NAC. Geometric, first order or textural (higher order) radiomic features were extracted from pre-NAC MRI and feature reduction was performed. Five radiomic features were added to other available information to build predictive models of complete response to NAC using three different classifiers (logistic regression, support vector machines regression and random forest) and exploring the whole set of possible feature selections.

Results: The study population consisted of 20 complete responders and 40 non-complete responders. Models including MRI radiomic features consistently showed better performance compared to combinations of other clinical, histological and radiological information. The AUC (ROC analysis) of predictors that did not include radiomic features reached up to 0.89, while all three classifiers gave AUC higher than 0.90 with the inclusion of radiomic information (range: 0.91-0.98).

Conclusions: Radiomic features extracted from 3T DCE-MRI consistently improved predictive models of complete response to neo-adjuvant chemotherapy. However, further investigation is necessary before this information can be used for clinical decision making.

Keywords: DCE; MRI; breast cancer; machine learning; medical imaging; neoadjuvant chemotherapy; radiomics.

Figure 1

LASSO variable selection process. (A) Values of the LASSO regression coefficient as a function of log (Lambda). (B) LOOCV deviance as a function of log (Lambda) and therefore of the number of selected features.

Figure 2

(A) Spearman correlation matrix between real variables and (B) Fisher’s p-values matrix between categorical variables.

Figure 3

Average AUC of the best 6 models as a function of the class of variables (G1, Rad, Hist, NoRad and All) and classifier (RF, random forest; SVR, Support Vector machines Regression; Logit, Logistic regression). Boxplots represent the median value, interquartile range and extremes.

Figure 4

Probability of each variable of being included in a high-performance model, estimated by the frequency with which the variable was selected in one of the 6 best models, as a function of the classifier.