Abstract
Deep learning models for medical imaging often struggle with data scarcity, particularly for rare pathologies. Traditional data augmentation techniques frequently fail to maintain anatomical realism or multiparametric consistency. This paper presents a two-stage generative framework designed to synthesize high-fidelity, multiparametric MRI scans (T1 arterial, T1 portal, and T2) for Massive Macrotrabecular Hepatocellular Carcinoma (MMHCC). By employing a mask-guided approach that incorporates anatomical constraints and edge detection, the method generates synthetic triplets that are perfectly co-registered. Trained on a limited dataset of 89 patients, the model produced 1,000 synthetic triplets with a Frechet Inception Distance (FID) score of 86.55, outperforming standard generative models and winning recognition at the 2021 French Society of Radiology data augmentation challenge.
Introduction
The efficacy of deep learning in medical imaging is often bottlenecked by the availability of high-quality, labeled datasets. This challenge is magnified in the study of rare diseases such as Massive Macrotrabecular Hepatocellular Carcinoma (MMHCC), a specific subtype of liver cancer. While data augmentation is a standard solution to expand training sets, medical images present unique constraints. Simple geometric transformations—like horizontal flipping—can inadvertently create anatomically incorrect "situs inversus" conditions. Furthermore, multiparametric MRI protocols require different sequences (e.g., T1 and T2) to be perfectly aligned, a task that standard Generative Adversarial Networks (GANs) often fail to achieve.
Methodology: A Two-Stage Generative Framework
To address these challenges, the researchers developed a "mask-guided" architecture that enforces anatomical realism through a two-stage process.
Stage 1: Anatomical Mask Generation
The first stage focuses on creating a structural blueprint for the synthetic image.
- Tumor Transformation: Existing tumor masks are subjected to random transformations, including zooming, rotation, and translation. To ensure clinical validity, a "liver intersection mask" is applied, constraining the synthetic tumor to remain strictly within the boundaries of the liver.
- Edge Detection: A Canny edge detector is applied to real MRI scans to extract "abdominal edges." These edges provide the global anatomical context.
- Input Mask Synthesis: The transformed tumor masks and abdominal edges are combined into a single input mask ($N_{msp}$), serving as the spatial guide for the subsequent generation stage.
Stage 2: Multiparametric Image Synthesis
The second stage utilizes the Pix2Pix architecture, a conditional GAN. To ensure consistency across different MRI phases, three parallel Pix2Pix networks are trained—one for each sequence (T1-FS arterial, T1-FS portal, and T2-weighted). Because all three networks receive the same structural mask ($N_{msp}$) as input, the generated images are inherently co-registered, mimicking the multi-phase acquisition of a single patient.
Experimental Setup and Results
The framework was trained using a dataset provided by the French Society of Radiology (SFR) for their 2021 data augmentation challenge. The dataset consisted of 267 multiparametric MRIs from 89 patients with MMHCC. From this base, the authors generated 1,000 synthetic triplets and their corresponding masks.
The performance was evaluated using the Frechet Inception Distance (FID) score, which quantifies the similarity between the distribution of synthetic and real images. The proposed method achieved an FID of 86.55, a significant improvement over baseline models:
- Pix2Pix (Baseline): 234.39
- CycleGAN: 210.32
Qualitative evaluation by three radiologists further confirmed the diversity and fidelity of the synthetic images, noting their realism across different sequences.
Discussion and Conclusion
The success of this approach lies in the use of "anatomical guardrails." By restricting the GAN's generative freedom through liver-constrained masks and edge-based anchors, the model avoids the common pitfalls of vanilla GANs, such as organ distortion or impossible tumor placement.
While the results are promising, the study's reliance on a small initial dataset (89 patients) suggests that further improvements in texture realism could be achieved with more diverse training data. Future work will focus on "downstream validation"—determining if the inclusion of these 1,000 synthetic cases improves the performance of segmentation and classification models in clinical settings. This research demonstrates that mask-guided synthesis is a viable path forward for overcoming data scarcity in rare disease research.
Summary updated