Prediction of pathologic complete response to neoadjuvant systemic therapy in triple negative breast cancer using deep learning on multiparametric MRI

Abstract

Triple-negative breast cancer (TNBC) is an aggressive subtype of breast cancer. Neoadjuvant systemic therapy (NAST) followed by surgery are currently standard of care for TNBC with 50-60% of patients achieving pathologic complete response (pCR). We investigated ability of deep learning (DL) on dynamic contrast enhanced (DCE) MRI and diffusion weighted imaging acquired early during NAST to predict TNBC patients' pCR status in the breast. During the development phase using the images of 130 TNBC patients, the DL model achieved areas under the receiver operating characteristic curves (AUCs) of 0.97 ± 0.04 and 0.82 ± 0.10 for the training and the validation, respectively. The model achieved an AUC of 0.86 ± 0.03 when evaluated in the independent testing group of 32 patients. In an additional prospective blinded testing group of 48 patients, the model achieved an AUC of 0.83 ± 0.02. These results demonstrated that DL based on multiparametric MRI can potentially differentiate TNBC patients with pCR or non-pCR in the breast early during NAST.

Figure 1

Patient selection. A group of 130 patients was used for model development through fivefold cross-validation. A group of 32 patients was reserved for retrospective independent testing. An extra prospective group of 48 patients was used for blinded testing.

Figure 2

The receiver operating characteristic (ROC) plots and corresponding areas under the curves (AUCs) of each fold for the training (A), validation (B), retrospective independent testing (C), and prospective blinded testing (D) groups. The averaged AUCs for each group were 0.97, 0.82, 0.86, and 0.83, respectively. The training and validation groups were used to develop the deep learning network and determine the optimal operating thresholds.

Figure 3

The image curation process. The tumors were cropped by placing tight bounding boxes based on the segmentations. The baseline tumor crops were first rescaled to the median size, then we used the same rescaling factor to resize the C4 tumor crops. Finally, all rescaled crops were zero-padded to the same sizes of 80 × 80 × 24 for the PEI, and 80 × 80 × 12 for the DWI. Such preprocessing approach permits maintaining the relative tumor shape across the patients, maintaining the relative tumor size across two scans, and avoiding large amounts of zero-padding, which can be more efficient for model training. The signal intensities were then normalized between 0 and 1.

Figure 4

The deep learning network. The framework of the feature extraction module, shown at left, consisted of four input channels corresponding to baseline positive enhancement integral (PEI), baseline diffusion-weighted imaging (DWI), PEI acquired after 4 cycles of doxorubicin/cyclophosphamide treatment (C4 PEI), and DWI acquired after 4 cycles of doxorubicin/cyclophosphamide treatment (C4 DWI). The detailed structure of the feature extraction module of one channel is shown at right. The extracted features of the four channels were concatenated and used for breast pathologic complete response prediction. Two dropout layers were added after the densely connected layers to prevent overfitting of the model. Conv3D: 3D convolutional layer, BN: batch normalization, ReLU: rectified linear unit, MaxPooling3D: 3D max pooling layer.