Author: Mengqi Shen, Meghna Trivedi, Tony J.C. Wang, Andy (Yuanguang) Xu, Yading Yuan 👨🔬
Affiliation: Columbia University Medical Center, Dept of Med Hematology & Oncology, Data Science Institute at Columbia University, Columbia University Irving Medical Center, Department of Radiation Oncology, Columbia University Irving Medical Center 🌍
Purpose: T2-weighted (T2w) images are critical for identifying pathological changes due to their superior contrast in differentiating tissue types. However, they often lack detailed anatomical resolution compared to T1-contrast (T1c) images, especially in sagittal and coronal views. This limitation arises from voxel anisotropy, where slice thickness exceeds in-plane resolution, and clinical compromises between contrast, acquisition time, and signal-to-noise ratio (SNR). This study aims to achieve precise anatomical detail in all imaging planes without compromising the diagnostic utility of T2w images.
Methods: A total of 295 brain metastasis cases from the BRATS challenge were included, selected based on standard anatomical resolution of T1c in all views and T2w with at least standard resolution in the axial view. An AI-based framework involving preprocessing, generative adversarial network (GAN) modeling, transfer learning, and volumetric reconstruction was proposed. Preprocessing included normalization, masking, and patch-based segmentation. The GAN was trained to translate imaging contrasts at the patch level using adversarial and reconstruction losses. The trained model was extended to lower-resolution views to ensure consistency across orientations. Volumetric reconstruction synthesized isotropic volumes, integrating outputs from multiple views for uniform resolution and precise anatomical representation.
Results: The reconstruction loss based on the L1-norm for axial views was 0.08. The Structural Similarity Index (SSIM) for sagittal and coronal views was 0.74 and 0.84, respectively. Sharpness metrics among tumor regions showed an improvement of 74.5%. Visual inspection by radiologists of the enhanced T2w and T1c across the three axial views demonstrated that anatomical details are well-preserved in all orientations. When comparing the enhanced T2w with the original T2w, tissue-level contrast accuracy was achieved with significantly improved anatomical details.
Conclusion: The framework effectively enhanced T2w anatomical detail across all orientations while preserving diagnostic accuracy, addressing limitations of voxel anisotropy and clinical imaging constraints.