Dynamic Contrast Enhanced MRI Mapping of Vascular Permeability for Evaluation of Breast Cancer Neoadjuvant Chemotherapy Response Using Deep Learning

Introduction: Dynamic contrast enhanced (DCE) MRI is a non-invasive imaging technique that has become a quantitative standard for assessing tumor microvascular permeability. Through the application of a pharmacokinetic (PK) model to a series of T1-weighed MR images acquired after an injection of a contrast agent, several vascular PK parameters can be quantitatively estimated. These parameters, including Ktrans, a measure of capillary permeability, have been widely implemented for assessing tumor biology as well as tumor therapeutic response. However, conventional PK modeling in DCE MRI is complex and time-consuming for dynamic scans with thousands of pixels per image [1]. In the present study, an image-to-image conditional generative adversarial network (cGAN) approach is developed as a surrogate to conventional PK modeling. The cGAN’s utility to map vascular permeability is validated in breast cancer patient DCE MRI data acquired before and after the first cycle of neoadjuvant chemotherapy (NACT). Moreover, the NACT induced early changes in vascular permeability has been correlated with pathological treatment response.

Materials and

Methods: An open-source collection of breast cancer patient DCE MRI data provided by The Cancer Imaging Archive (TCIA) was used in the current study for training, testing, and validating the proposed cGAN approach (n = 2400 slices) [2,3]. The TCIA collection consists of 10 breast cancer patients who were each subject to two DCE MRI imaging exams, pre-NACT (visit 1, V1) and after the first cycle of NACT (V2). The treatment response of each patient to NACT (complete response or non-complete response) was determined by pathologic analyses following the last cycle of NACT and provided by the TCIA. PK modeling was performed using the Extended Tofts Model (ETM) to generate target Ktrans maps. For cGAN training, a generator network was trained to translate input DCE MRI and R10 maps to synthetic Ktrans maps. Simultaneously, a discriminator network was trained to distinguish between the generator synthetic Ktrans maps and real ETM Ktrans maps, conditioned on input DCE MRI and R10 maps. A custom patch-based adversarial loss incorporating an L1-loss component was used to update both generator and discriminator network parameters until resulting Ktrans maps produced by the generator were indistinguishable from real ETM Ktrans maps (Figure 1). 10-fold cross validation was conducted by excluding one patient at a time from the training process for testing of the cGAN. Linear regression was applied to provide statistical correlation and significance. NACT induced early changes in vascular permeability parameter Ktrans was correlated with pathological treatment response leveraging univariate logistic regression (ULR) models.

Results, Conclusions, and Discussions: The proposed image-to-image cGAN can predict vascular permeability in TCIA datasets in 2.5 seconds on average, significantly lowering the average computation time over 1000-fold relative to the ETM (150 minutes on average). As shown in Figure 2, representative DCE MRI and ETM Ktrans maps revealed high vascular permeability of breast cancer prior to the start of treatment, V1, for both a complete responder (Figure 2a) and a non-complete responder (Figure 2b). Notably, the complete responder had markedly decreased Ktrans at V2, whereas the non-complete responder retained elevated vascular permeability. Concurring with the ETM, the cGAN recapitulated these vascular permeability trends in both the complete and non-complete responders (Figure 2a and 2b). Quantitative analyses showed that there was a strong correlation between the cGAN and ETM measurements of intratumoral pixel-wise Ktrans (R2 > 0.94; p < 0.0001) and slice-wise mean Ktrans (R2 > 0.96; p < 0.0001). We further explored the prognostic value of DCE MRI for predicting breast cancer NACT response. As presented in Figure 3, scatter plots of tumor mean Ktrans at V1 and V2 revealed a more prominent decrease in Ktrans for complete responders. Clearly, there was an excellent agreement between the ETM and cGAN of tumor mean Ktrans changes post-NACT. Statistically, ULR c-statistic values revealed V2 Ktrans (c-statistic = 0.952) and percentage Ktrans change (c-statistic = 1.00) provide excellent early prediction of pathologic response for both the ETM and cGAN. In conclusion, we have developed an image-to-image cGAN approach to accurately map vascular permeability in breast cancer and its change in response to NACT. We anticipate that the cGAN approach developed in this study may serve as a fast and accurate tool for MR imaging of breast cancer diagnosis and prognosis, potentially improving treatment planning and patient outcomes.

Acknowledgements (Optional):

References:
[1] Arledge CA et al. Cancers (Basel). 2023.
[2] Huang W et al. The Cancer Imaging Archive. 2014.
[3] Huang W et al. Translational Oncology. 2014.