Assimilating dual decoder precision with adversarial training for realistic medical data synthesis
Year: 2026
Type: article
Abstract: This study investigates the challenge of limited high-quality annotated medical imaging data by proposing an advanced Variational Autoencoder with Wasserstein Generative Adversarial Network (VAE-WGAN) framework, termed Dual-Decoder Skip VAE-WGAN (D2S-VAE-WGAN), designed to generate realistic multi-modal MRI data. The proposed architecture incorporates skip connections to enhance detail preservation and employs a dual-decoder approach—one dedicated to reconstructing multi-modal images (FLAIR, T1, T1ce, T2) and another to generating segmentation masks. Additional loss functions, including perceptual loss, dice loss, and boundary loss, are integrated to ensure both fidelity and structural precision. This framework enhances dataset diversity and enables resource-efficient processing of 3D volumes, optimized for environments with 16GB VRAM and constrained computational sessions, such as Kaggle’s 12-hour limits. Evaluation using a benchmark dataset produced 250 synthetic samples, achieving a mean Structural Similarity Index (SSIM) of 0.7909, Peak Signal-to-Noise Ratio (PSNR) of 22.6535 dB, and an average Dice score of 0.7409. Sub-region Dice scores include 0.7002 for necrotic core (NCR), 0.8008 for enhancing tumor (ET), 0.7216 for edema (ED), 0.8411 for whole tumor (WT), and 0.8798 for tumor core (TC). This research presents a novel, reproducible pipeline that significantly contributes to generative modelling in medical imaging, providing a scalable solution for data-constrained settings.
Source: Systems and Soft Computing
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,870
OpenAlex ID: W7153529237