This study retrospectively collected the breast MRI images from 97 patients who underwent MRI examinations between 2019 and 2021 at Xi’an International Medical Center Hospital, and patients were randomly split into a training set (n = 45), a validation set (n = 31) and a test set (n = 21), respectively. The inclusion criteria for the MRI images were as follows: I) image acquisition at 3.0T magnetic field; II) excellent image quality with no motion artifacts. The exclusion criteria were as follows: I) images with an incomplete sequence; II) incomplete clinical data; and III) a history of surgical resection. This research was approved by the Ethics Committee of the Xi’an International Medical Center Hospital and was conducted according to the principles of the Declaration of Helsinki.

Image acquisition was performed using the MAGNETOM Prisma 3.0T MRI. The imaging protocol mainly included T1-weighted (T1w), T2-weighted (T2w), DWI, and T1w sequences after contrast administration. Here, the Gd-DTPA (0.1 mmol per kilogram body weight) was intravenously injected at an injection rate of 2mL/s, and then the same amount of normal saline was injected. Furthermore, the scan parameters were as follows: Axial T1w: 176 slices, slice thickness = 1.0 mm; FOV =384×384 mm; TR/TE, 5.66/2.46 ms; matrix = 384×384. Axial T2w: 160 slices, slice thickness = 0.9mm; FOV =360×230 mm; TR/TE, 2000/220ms; matrix = 400×256. Axial gadolinium-enhanced T1W: 176 slices, slice thickness = 1.0 mm; FOV = 384×384 mm; TR/TE, 4.66/1.62 ms; matrix = 384×384. Axial DWI: 35 slices, slice thickness = 4.0 mm; TR/TE, 6100/65ms; FOV = 168×340 mm, b = 0 s/mm² and 1,000 s/mm²; EPI factor = 84; matrix = 84 ×170; bandwidth = 2262 Hz/pixel.

We used the following steps for data preprocessing. Firstly, the original T1WI and FP-Dyn images in DICOM format were converted to PNG format by using the MicroDicom Viewing software (http://www.microdicom.com/). Then, TIWI images were subtracted from the original FP-Dyn images, and applied a threshold to obtain the contrast agent enhancing areas. Furthermore, using the FP-Dyn images, the edge detail label was obtained through further Canny edge detection. Moreover, the pixel values for each image were scaled into the range [0,1] by using the min-max scaling method. Finally, we expanded the training datasets through the data augmentation strategies, which included rotating, cropping, and mirroring them to improve the performance and robustness of the model (23, 24).

The enhance border lifelike synthesize (EDLS) model consists of two components: a segmentation network (stage I) and a synthesis network (stage II). The two networks have a similar network structure, including a generator using the U-Net network and a discriminator based on the convolutional neural network, respectively. First, the segmentation network was constructed to segment the potentially contrasted agent enhancing areas of T1WI images. Then, the synthesis network was used to produce the FP-Dyn sequence images from the input T1WI images and the segmentation information. Here, the segmentation information was used as additional information to supervise the contrasted agent enhancing areas in the synthetic images ( Figure 1 ).

Furthermore, the Dice loss function was used to train the segmentation network in order to tackle the class imbalance problem in the enhanced area and the non-enhanced area. As shown in formula 1.

Where, Ω indicated the set of pixels in the whole image, s represented the pixel, and p(s) represented the predicted probability for s category by the segmentation network, y(s) represented the true category label of pixel s, and ε denoted the smoothing coefficient.

In the synthesis network, the edge loss function and depth supervision strategy were introduced to solve the problem of blurring of edges and details in the synthesized image. The detailed information was displayed in formula 2-6.

The G and D represented the generator and discriminator respectively. x represented the original T1WI sequence images, x_seg was the segmented image corresponding to x, y was the original FP-Dyn sequence image, y_edge represented the y image edge information map obtained by Canny operator detection, l_cGAN represented the generation countermeasure loss function, and the loss function l_layer_L1 was introduced to further ensure the similarity between the synthesized image and the original image. In addition, to solve the problem of lacking edge detail information in the synthesized FP-Dyn images, we added the edge detail loss function l_layer_L2 into the model. Furthermore, the depth supervision network l_DS was added into the model for stable convergence. And, the λ₁ and λ₂ were weighted coefficients and assigned values of 1 and 6 according to the experimental experience, respectively.

The EDLS model training task sought to learn the mapping specifically between T1WI sequence images and FP-Dyn sequence images. Here, the EDLS model was trained on a training set including 3996 pairs of T1WI-FP-Dyn images. We optimized the loss functions by using the Adam optimizer during the training stage, and the model parameters were updated for each training iteration until the model converged. In addition, we implemented all the models with Python 3.6. The TensorFlow framework was used for model construction and evaluation. The software and hardware included CUDA9.0, cuDNN7.6.5, and a Linux server with 2 NVIDIA GTX 1080Ti GPUs.

Following the completion of the model training, the EDLS model was used to synthesize the FP-Dyn images and sequences based on 1226 T1WI images and 25 T1WI sequences in the validation set, respectively. Furthermore, we also synthesized the FP-Dyn sequences using the EDLS model for each patient in the test set.

We compared the performance of the EDLS model to conventional GAN models, including CycleGAN (25), DC2Anet (26), MR-GAN (27), and Pix2Pix (28). Conventional models were trained and evaluated using the same data set as the EDLS model.

The model performance was evaluated by using the quantitative index of peak signal-to-noise ratio (PSNR), structural similarity (SSIM), mean square error (MSE) and mean absolute error (MAE) (29). A lower MSE and MAE, or a higher PSNR and SSIM, indicated a better model performance for image synthesis (30).

The test images, including the synthesized FP-Dyn images and the original FP-Dyn images, were randomly put together. Three radiologists independently discriminated between the synthesized and original FP-Dyn images. In addition, with reference to the original image, the satisfaction of the synthesized images was evaluated by the radiologists using a five-point Likert scale (1 disagree strongly, 2 disagree, 3 no response, 4 agree, and 5 agree strongly). Further, we dichotomized the evaluated results. Scores equal to or greater than 4 were defined as “satisfied”. Scores of 1 to 3 were defined as “dissatisfied”.

Three reading modes, including reading mode (a), reading mode (b), and reading mode (c), were designed to evaluate the diagnostic value of the synthesized FP-Dyn sequence. And the reading mode (a) included T1WI, T2WI, DWI, and the synthesized FP-Dyn sequence. The reading mode (b) included the T1WI, T2WI, DWI and the originally scanned FP-Dyn sequence. The reading mode (c) included the T1WI, T2WI, DWI and the total phases of scanned DCE sequences. Then, three experienced breast radiologists independently diagnosed under each of the three modes, respectively. There was a minimum of one month between each of the reading modes. The breast lesions were diagnosed according to the BI-RADS classification. Here, a BI-RADS score greater than 3 was considered positive, while 3 or less was considered negative. We compared the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) among the afore-mentioned three reading modes, using the pathology results as the gold standard.

In this study, continuous variables were presented as mean and standard deviation, and categorical variables were presented as percentages. The Analysis of Variance (ANOVA) test was used to analyze the performance differences between our model and the conventional GAN models. In addition, the diagnostic consistency among three different reading modes was calculated from the Kappa test. Furthermore, the Kendall test was used to evaluate the consistency of the diagnoses among the radiologists. A two-side P-value less than 0.05 was considered statistically significant. The Holm method was used to adjust the P value between multiple group comparisons. All the statistical tests were performed in R (version 3.6.3, https://www.rproject.org) software.

We quantitatively compared the 1226 FP-Dyn images which were respectively synthesized by the EDLS model and the conventional GAN models. Here, the results which were measured by the MAE, MSE, PSRN, and SSIM were summarized in Figure 2 . It was easy to see that the performance of the EDLS model was significantly superior to other conventional GAN models. Compared with the Pix2Pix model, the EDLS model produced improvements of more than 2% in SSIM, which played an important role in improving the quality of the synthesized images. Simultaneously, the PSNR had also been significantly improved (P<0.001). In addition, we found that the EDLS model had the lowest MAE and MSE (P<0.001).

The results of PSNR, SSIM, MAE and MSE from the synthesized FP-Dyn sequences in 25 patients were shown in Supplementary Figure 1 . The experimental results suggested that the EDLS model had the highest PSNR and SSIM (P<0.001), and had the lowest MAE and MSE (P<0.001), compared with the conventional models ( Supplementary Figure 1 ).

The 1226 synthesized FP-Dyn images and 1226 original FP-Dyn images were randomly put together. As seen in Table 1 , one radiologist correctly identified images with 52.00% accuracy of 2452 images, and the precision rate for correctly identifying the original FP-Dyn image was 51.20% of 1226 images. In addition, each of the other two radiologists correctly identified images with 56.65% and 53.67% accuracy, and 54.35% and 52.71% precision.

Qualitative metrics for image quality evaluation were shown in Table 2 . And there were no significant differences in subjective image quality scores among the three radiologists. Compared with the original FP-Dyn images, more than 99% of the synthesized images in the shape consistency were good, and the scores were 4 or 5. In addition, the contrast enhancement for synthesized images was also given high scores by the three radiologists. For example, more than 91% of synthesized FP-Dyn images got good scores of 4 or 5 points for great vessels and heart enhancement. And more than 75% of synthesized FP-Dyn images in gland enhancement got good scores of 4 or 5 points. Furthermore, from the evaluation results, we found that the synthesized images were effective for the suppression of motion artifacts.

In addition, Figure 3 showed an example of the input T1WI images, synthesized FP-Dyn images, original FP-Dyn images, and absolute error images. The absolute error images showed the absolute error between the synthesized FP-Dyn image and the original FP-Dyn image. We could clearly see that the EDLS model had been learning to identify tissues that had similar signal values in T1WI sequences, but had high signals in FP-Dyn sequences, such as tumor tissue, the heart, and great blood vessels. And the FP-Dyn image synthesized by the EDLS model was highly consistent with the original FP-Dyn image and had a similar enhancement in the lesion. However, the images synthesized by the CycleGAN model failed to effectively enhance the lesions. In addition, there were high background pixel errors from the FP-Dyn images which were synthesized by other models, compared with their original images.

We further evaluated the reconstruction error of the synthesized FP-Dyn sequence on the validation set. From Figure 4 , we could see that the synthesized FP-Dyn sequence images had a small reconstruction error compared with the original FP-Dyn sequence images. The average MAE on all sequences was 0.027 with a standard deviation of 0.004, and the average MSE on all sequences was 0.003 with a standard deviation of 0.001. In addition, some outliers of reconstruction errors were detected in the statistical boxplot. And we found there were some gland enhancements in the synthesized FP-Dyn images lower than in the original FP-Dyn images. It might be the major reason for the outliers of reconstruction errors in the boxplot.

In addition, we assessed the correlation between the synthesized FP-Dyn sequences and the original FP-Dyn sequences for each patient in the validation set, then an overall average for sequence correlations was calculated. Here, we found that there were high positive correlations between the synthesized FP-Dyn sequences and the original FP-Dyn sequences. The overall average of the correlations between the synthesized FP-Dyn sequences and the original FP-Dyn sequences was r = 0.927 ± 0.311 (95% CI: 0.927 to 0.928, P <0.001), ( Supplementary Figure 2 ).

We collected a test set of 21 patients who underwent breast MRI examinations and with pathological examination results. Of 11 cases that were diagnosed as benign tumors, 1 case of chronic mastitis, 2 cases of benign epithelial hyperplasia, 2 cases of adenoids, 1 case of fibroadenoma, 1 case of fibroadenoma with adenopathy, 2 cases of fibroadenoma with benign epithelial hyperplasia, and 2 cases of invasive ductal carcinoma without cancer cells in pathology after radiotherapy. The remaining 10 patients had malignant lesions. Of 10 cases that were diagnosed as malignant lesions, 7 cases were non-specific invasive ductal carcinoma, 1 case was non-specific invasive ductal carcinoma with necrosis, and 2 cases were metastatic breast ductal carcinoma. Here, the patients were respectively diagnosed by three experienced breast radiologists using the three-reading mode. The three radiologists had high diagnosis consistency as measured by the Kappa test (Kappa = 0.688, P<0.001).

Based on pathology results, we counted the true-positive (TP), false-negative (FN), false-positive (FP), and true-negative (TN) values diagnosed by the three models, and then the sensitivity, specificity, PPV, and NPV values for each reading mode were then calculated, which were listed in Table 3 . The sensitivity, specificity, PPV, and NPV of the three reading modes for breast cancer diagnosis were as follows. The reading mode (a) was 100%, 72.73%, 76.92%, and 100%, respectively. The reading mode (b) was 100%, 63.64%, 71.43%, 100%, respectively. The reading mode (c) was 100%, 72.73%, 76.92%, and 100%, respectively ( Table 3 ).

Importantly, the reading mode (a) and reading mode (b) had the same diagnosis in 20 cases, accounting for 95.24%. The reading mode (a) and reading mode (c) had the same diagnosis results in 19 cases, which accounted for 90.48%. Each of the breast cancer cases both obtained a diagnosis with high BI-RADS scores (4 or 5) by three reading modes, respectively.