A Comprehensive Framework Analysis of Cycle GAN-Based Modality Translation: Enhancing Brain Tumor Diagnostics from FLAIR to T2w

Md Abubakkar

Abstract:

Brain tumors remain a major healthcare challenge
due to their complexity, high mortality rates, and profound im-
pact on patients’ lives, making accurate diagnosis and treatment
crucial. Multi-modal MRI scans, particularly FLAIR and T2-
weighted (T2w) images, offer complementary information about
tumor structure and progression. However, real-world clinical
settings often face the challenge of missing imaging modalities,
limiting comprehensive assessments. To address this issue, we
propose a CycleGAN-based framework for translating between
FLAIR and T2w MRI scans using the BraTS dataset from the
Medical Decathlon Challenge. This dataset includes 3D MRI
scans with segmentation masks outlining key tumor regions
such as edema, non-enhancing tumor, and enhancing tumor.
Our framework uses U-Net-based generators and PatchGAN
discriminators, optimized with multiple loss functions, including
adversarial, cycle consistency, structural similarity index (SSIM),
and pixel-wise losses. These ensure that the generated images
are both anatomically accurate and visually realistic. We apply
preprocessing steps like intensity normalization, background
removal, and data augmentation to maintain structural details
and enhance training stability. Our quantitative evaluation shows
promising results, achieving SSIM scores of 0.8226 for T2w
and 0.7767 for FLAIR. Qualitative analysis further highlights
improved tumor visibility and clearer anatomical structures,
particularly around tumor boundaries. By addressing the chal-
lenge of incomplete imaging datasets, our method not only
enhances data availability for tasks like tumor segmentation
but also supports more comprehensive diagnostic workflows.
This approach represents a step forward in advancing precision
medicine for brain tumor analysis through multi-modal MRI
synthesis.