Author: Amir Abdollahi, Oliver Jäkel, Maxmillian Knoll, Rakshana Murugan, Adithya Raman, Patrick Salome 👨🔬
Affiliation: UKHD & DKFZ, Department of Medical Physics in Radiation Oncology, German Cancer Research Center (DKFZ), German Cancer Research Centre(DKFZ), DKFZ, MGH 🌍
Purpose:
Missing MRI sequences, due to technical issues in data handling or clinical constraints like contrast agent intolerance, limit the use of medical imaging datasets in computational analysis and research. We aimed to develop a framework that addresses this challenge by synthetically generating missing MR sequences.
Materials/Methods:
The study utilized two datasets: BraTS2021, a public dataset containing standardized mpMRI sequences (T1w, T1wce, T2w, FLAIR), and a private dataset of 101 recurrent high-grade glioma patients treated with carbon-ion radiotherapy (2009-2018). We built a pipeline around pix2pix deep learning (DL) generative adversarial networks (GAN) models, optimizing hyperparameters to generate synthetic contrast-enhanced T1w (T1wce) images while evaluating three configurations: single-sequence (1-to-1), dual-sequence (2-to-1), and triple-sequence (3-to-1) inputs. Synthetic image quality was assessed through peak signal-to-noise ratio (PSNR) and structural similarity index metrics (SSIM). Quantitative radiomics features were extracted from both real and synthetic images, and these features were used in subsequent analyses, including correlation analysis. Lastly, survival analysis using Cox Proportional Hazard models was performed, comparing the performance of the synthetic data approach against the traditional mean imputation method.
Results:
The 3-to-1 region-based model achieved the best performance with a PSNR of 30.25 and an SSIM of 0.98, representing improvements of 14.90% and 7.15%, respectively, compared to the image-based approach. Survival prediction using synthetic data achieved comparable performance to models using original data (C-Index: 0.6819 [0.6304, 0.7337] vs 0.6813 [0.6271, 0.7356]). Finally, in radiomics-based survival modelling, the use of the synthetically generated images for missing data outperformed traditional imputation methods (C-Index: 0.63 [0.5525, 0.7075] vs 0.5873 [0.5380, 0.6366]).
Conclusion:
Deep learning-based synthetic generation shows promising results for addressing missing MRI data. Region-focused and multi-sequence approaches outperform whole-image generation, suggesting future pipelines should combine specialized regional models to reconstruct complete images.