Synthesizing the First Phase of Dynamic Sequences of Breast MRI for Enhanced Lesion Identification

Pingping Wang, Pin Nie, Yanli Dang, Lifang Wang, Kaiguo Zhu, Hongyu Wang, Jiawei Wang, Rumei Liu, Jialiang Ren, Jun Feng, Haiming Fan, Jun Yu, Baoying ChenFrontiers in Oncology 2021 · DOI

TL;DR

This paper introduces a novel deep learning model, EDLS, capable of synthesizing dynamic contrast-enhanced (DCE) MRI sequences for breast cancer detection. By generating these crucial "first phase dynamic" (FP-Dyn) images from standard T1-weighted MRI scans, the method aims to reduce examination time, cost, and the need for contrast agents, without compromising diagnostic accuracy.

What Problem Does This Paper Address?

Breast cancer screening and diagnosis heavily rely on MRI, particularly dynamic contrast-enhanced (DCE) sequences, which highlight lesion characteristics. However, traditional full MRI protocols are time-consuming, expensive, and carry risks associated with contrast agents (e.g., allergic reactions). Unenhanced abbreviated MRI (u-AB-MRI) protocols offer a solution by omitting contrast agents and shortening scan times, but this often leads to a loss of diagnostic information due to the absence of DCE sequences. This paper addresses the critical need for a method to generate these essential DCE sequences without requiring actual contrast administration or extended scanning, thereby bridging the gap between comprehensive but lengthy protocols and abbreviated but potentially less informative ones.

Methods

The core of this work is the Enhance Border Lifelike Synthesize (EDLS) model, a novel deep learning architecture designed for image-to-image translation. EDLS is a two-stage approach:

  • Stage I (Segmentation Network): A generator (G1) and discriminator (D1) pair is used to segment potentially contrast-enhancing areas from T1-weighted (T1WI) images. This stage aims to isolate regions of interest that would typically show enhancement after contrast injection.
  • Stage II (Synthesis Network): A second generator (G2) and discriminator (D2) pair takes the original T1WI images and the segmented enhancement information from Stage I to synthesize the first phase of dynamic (FP-Dyn) sequences.

Key innovations in EDLS include:

  • Edge Loss Function: To ensure the synthesized images retain fine details and sharp edges, an edge loss is incorporated.
  • Depth Supervision Strategy: This is used to improve the stability and accuracy of the synthesis process.
  • Feedback Mechanism: The architecture includes a feedback loop to refine the synthesized output.

The model is trained using a combination of loss functions, including a Dice loss for segmentation, a GAN loss (lcGAN), an L1 loss (llayer_L1) for pixel-wise similarity, and an L2 loss (llayer_L2) specifically for edge details.

Datasets

The study utilized a dataset of breast MRI images from 97 patients. These were divided into:

  • Training set: 45 patients
  • Validation set: 31 patients
  • Test set: 21 patients

The MRI protocols included T1-weighted (T1w), T2-weighted (T2w), diffusion-weighted imaging (DWI), and T1w sequences after contrast administration (FP-Dyn). The paper does not mention the introduction of new benchmark datasets.

Key Insights and Ideas

  • Synthesizing Crucial Information: The most significant insight is that it's possible to generate clinically relevant FP-Dyn MRI sequences from non-contrast T1WI images using deep learning. This bypasses the need for contrast agents and reduces scan time.
  • EDLS Outperforms Conventional GANs: The proposed EDLS model demonstrated superior performance compared to established GAN architectures like CycleGAN, DC2Anet, MR-GAN, and Pix2Pix in quantitative metrics. This suggests that the specific architectural choices and loss functions in EDLS are well-suited for this complex medical imaging synthesis task.
  • Visual Similarity is High: Radiologists found it difficult to distinguish between synthesized and original FP-Dyn images, indicating a high degree of visual fidelity. This is crucial for clinical acceptance.
  • Diagnostic Value Maintained: Critically, the synthesized FP-Dyn sequences, when combined with unenhanced MRI sequences, provided diagnostic performance (sensitivity, specificity, PPV, NPV) comparable to full DCE-MRI protocols. This is the ultimate validation for its clinical utility.
  • Reduced Motion Artifacts: Synthesized images showed fewer motion artifacts compared to actual scanned images, potentially improving image quality.

Key Results

  • Quantitative Performance: EDLS achieved significantly higher PSNR and SSIM, and lower MAE and MSE compared to conventional GAN models. For example, EDLS showed improvements of over 2% in SSIM compared to Pix2Pix.
  • Visual Evaluation: Radiologists had an accuracy of around 52-56% in distinguishing between synthesized and original FP-Dyn images, suggesting they are visually very similar.
  • Subjective Quality: Radiologists rated the synthesized images highly for shape consistency, enhancement, and artifact suppression, with no significant differences between readers.
  • Diagnostic Value: In the test set, using T1WI, T2WI, DWI, and synthesized FP-Dyn sequences (reading mode a) resulted in sensitivity of 100%, specificity of 72.73%, PPV of 76.92%, and NPV of 100%. These values were comparable to or better than using only unenhanced sequences (reading mode b) and equivalent to full DCE-MRI (reading mode c).
  • Reconstruction Error: The average MAE for synthesized FP-Dyn sequences on the validation set was 0.027 ± 0.004, and average MSE was 0.003 ± 0.001, indicating low reconstruction error.

Limitations and Future Work

  • Dataset Size: The study used a relatively small dataset (97 patients) for training and testing, which may limit the generalizability of the model across diverse patient populations and imaging centers.
  • Automatic Thresholding: An automatic threshold segmentation algorithm was used to obtain lesion labels, which might not be optimal for all cases. Manual segmentation could improve accuracy.
  • Clinical Application Verification: The authors emphasize the need for further prospective, multi-center studies to fully validate the clinical application value of the EDLS model.
  • Future Directions:
    • Expand the dataset and retrain the model using multi-parametric sequences (T1WI, T2WI, DWI) to synthesize DCE images.
  • Explore manual lesion segmentation for improved accuracy.
  • Conduct multi-center, prospective studies to verify clinical utility.
  • Integrate the model into existing medical image processing platforms for easier deployment.

Summary updated

Abstract

The EDLS model could synthesize the realistic FP-Dyn sequence to supplement the lack of enhanced images. Compared with full MRI examinations, it thus provides a new approach for reducing examination time and cost, and avoids the use of contrast agents without influencing diagnostic accuracy.
BibTeX
@inproceedings{wang2021synthesizing,
  author = {Pingping Wang and Pin Nie and Yanli Dang and Lifang Wang and Kaiguo Zhu and Hongyu Wang and Jiawei Wang and Rumei Liu and Jialiang Ren and Jun Feng and Haiming Fan and Jun Yu and Baoying Chen},
  title = {{Synthesizing the First Phase of Dynamic Sequences of Breast MRI for Enhanced Lesion Identification}},
  booktitle = {Frontiers in Oncology},
  year = {2021},
  doi = {10.3389/fonc.2021.792516},
}
Figure