Cross-Modality MRI Translation: Unpaired and Bidirectional Learning with a Disentangled GAN with Gabor-Edge CNNs

Multi-modality MRI imaging is crucial for accurate disease diagnosis and detection, providing complementary information across modalities. However, acquiring multiple modalities is often constrained by patient discomfort, economic limitations, and extended scanning times. To address this, this work proposes a novel encoder-decoder GAN for unsupervised multi-modality MRI translation. Existing models face challenges in disentangling style (modality-dependent) and content (modality-independent) features, which is addressed by designing multiple encoders to capture disentangled features. Style-specific latent features are extracted using a conventional encoder, while content features are modelled through a Gabor / Edge CNN, leveraging spatial and frequency domain information. To overcome the limitations of U-shaped networks in capturing global-local features, this work integrated weighted dilated convolutions, emphasizing local features while down-weighting global ones for improved multi-scale contextual learning. The generator is conditioned on modality labels, enabling translation from a single source modality to multiple target modalities while simultaneously synthesizing segmented tumor regions. To ensure accurate tumor reconstruction, this work incorporates Dice loss for region-based optimization, ensuring precise synthesis within the translated modalities. The discriminator enhances perceptual quality via adversarial learning, producing high-fidelity translations that preserve anatomical structures. The proposed model is trained on the BraTS2018 dataset, which includes multiple MRI modalities, and is evaluated against state-of-the-art techniques using PSNR, SSIM, and computational metrics. Additionally, t-SNE is employed to visualize the features both in the image space—across input and translated MRI modalities—and in the latent representation space. The evaluation metrics and t-SNE visualizations demonstrate the superior performance of the proposed model compared to existing state-of-the-art approaches. Our approach addresses critical challenges in multi-modality MRI synthesis, offering a robust framework for accurate and efficient medical imaging translations.