RETURN TO ABSTRACT LISTING


Objective:
Neoadjuvant chemotherapy (NAC) response is routinely assessed using breast MRI due to its high sensitivity for detecting breast cancer. However, there are concerns regarding the unknown clinical significance of gadolinium deposition. In the setting of NAC assessment where MRIs are performed several times in a short period of time, there is potential benefit for a noncontrast alternative method to assess NAC response. The purpose of this study is to develop simulated postcontrast breast images from precontrast MRI sequences using a generative adversarial network (GAN) in patients undergoing NAC to assess treatment response.

Materials and Methods:
An IRB-approved study was performed. Breast MRI cases from patients undergoing NAC were collected from publicly available ISPY-2 trial (ACRIN 6698). Precontrast sequences that were used include T1 FS, T2 FS, and DWI/ADC. Preprocessing steps were performed including segmentation, coregistration of all images and Z-score intensity normalization. Briefly, a generative adversarial network (GAN) was used to model domain relationships. The first domain was contrast-enhanced MRI images. The second domain was precontrast images (T1,T2, ADC, dwib0, dwib100, dwib800, dwib1000). A cycle GAN technique was used to optimize similarity between simulated and true postcontrast images. One hundred pretreatment MRIs in 100 patients were randomly selected for training, and 10 post-NAC MRIs in 10 patients with pathologic complete response (pCR) and 10 post NAC MRIs in 10 patients with non-pCR were randomly selected for testing. Quantitative assessments were used to determine similarity between simulated and true postcontrast images including the degree of overlap and maximum diameter of tumor measurements.

Results:
In 7 of 10 MRIs with non-pCR, the simulated postcontrast images demonstrated significant overlap with the true contrast images (dice coefficient, 0.77 ± 0.22). The mean maximum tumor sizes between simulated and true postcontrast images were not significantly different (mean difference 2.5 mm, p = 0.4). In 3 of 10 MRIs with non-pCR, the network yielded false negative simulation. In 6 of 10 MRIs with pCR, the simulation showed no enhancement confirmed by internal controls. In 1 of 10 MRIs with pCR, the simulation showed false positive 0.9 cm enhancement. In 3 of 10 MRIs with pCR, the simulation showed no enhancement but was considered indeterminate.

Conclusion:
It is feasible to use simulated MRIs for NAC assessment, which could decrease gadolinium exposure with potential clinical utility in assessing partial or no response to NAC. Further work is needed to limit false negative cases that can mimic pCR.