🚧 Under Construction 🚧 Stay tuned! Updates coming soon...
The synthesis of T1-weighted contrast-enhanced (T1CE) MR images is essential for accurate brain pathology diagnosis while mitigating the health risks from gadolinium-based agents. This thesis introduces a novel two-stage deep learning approach using the BraTS2021 dataset, employing a modified U-Net model for both feature extraction and image synthesis. Our approach, incorporating constrained contrastive learning (CCL) in the full decoder, demonstrated improvements over the baseline. P-values from hypothesis testing with a 5% error rate suggest that the CCL model is significantly better than the baseline in terms of PSNR, SSIM, LPIPS (AlexNet) and LPIPS (VGG). These results highlight the potential of our methodology to significantly enhance diagnostic accuracy and patient safety in clinical settings.
![synth-model](https://private-user-images.githubusercontent.com/54392924/328063163-f38983f4-b5a3-4da8-aa6e-0b29051590c6.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.m8-8DyPOIsSggzdG_8_ZmYwS2pzW-yuQHCpI9WFACB0)