Machine Learning Models and Multiparametric Magnetic Resonance Imaging for the Prediction of Pathologic Response to Neoadjuvant Chemotherapy in Breast Cancer

Abstract

Background: Most breast cancer (BC) patients fail to achieve pathological complete response (pCR) after neoadjuvant chemotherapy (NAC). The aim of this study was to evaluate whether imaging features (perfusion/diffusion imaging biomarkers + radiomic features) extracted from pre-treatment multiparametric (mp)MRIs were able to predict, alone or in combination with clinical data, pCR to NAC.

Methods: Patients with stage II-III BC receiving NAC and undergoing breast mpMRI were retrospectively evaluated. Imaging features were extracted from mpMRIs performed before NAC. Three different machine learning models based on imaging features, clinical data or imaging features + clinical data were trained to predict pCR. Confusion matrices and performance metrics were obtained to assess model performance. Statistical analyses were conducted to evaluate differences between responders and non-responders.

Results: Fifty-eight patients (median [range] age, 52 [45-58] years) were included, of whom 12 showed pCR. The combined model improved pCR prediction compared to clinical and imaging models, yielding 91.5% of accuracy with no false positive cases and only 17% false negative results. Changes in different parameters between responders and non-responders suggested a possible increase in vascularity and reduced tumour heterogeneity in patients with pCR, with the percentile 25th of time-to-peak (TTP), a classical perfusion parameter, being able to discriminate both groups in a 75% of the cases.

Conclusions: A combination of mpMRI-derived imaging features and clinical variables was able to successfully predict pCR to NAC. Specific patient profiles according to tumour vascularity and heterogeneity might explain pCR differences, where TTP could emerge as a putative surrogate marker for pCR.

Keywords: imaging biomarkers; machine learning; multiparametric MRI; radiomics.

Figure 1

Box-and-whisker plots comparing imaging feature values between patients showing pathological complete response and non-responders.

Figure 2

Correlation matrix representing selected imaging features. Colour scale on the left side indicates degree of correlation (from −1 to 1 and from blue to red).

Figure 3

Accuracy values (left) for each of the tested classifiers and confusion matrices (right) corresponding to models trained with (A) imaging features (diffusion/perfusion MRI parameters + radiomic features), (B) clinical variables, and (C) imaging data + clinical variables. Classifiers achieving the highest accuracy values for each model are highlighted in green. AdaBoost = Adaptive Boosting; DT = Decision Tree; FN = false negative; FP = false positive; GBoost = Gradient Boosting; GNB = Gaussian Naive Bayes; K-NN = K-Nearest Neighbour; LDA = Linear Discriminant Analysis; LG = Logistic Regression; MLP = Multi-Layer Perceptron; QDA = Quadratic Discriminant Analysis; TN = true negative; TP = true positive.