Post-Processing Methods for Improving Accuracy in MRI Inpainting

TL;DR

This work addresses the limited applicability of automated MRI tools in the presence of large brain lesions by proposing a modular post-processing pipeline for 3D brain inpainting. It combines predictions from top BraTS inpainting methods with lightweight post-processing steps—voxel-wise averaging, median/Gaussian filtering, histogram matching, and a U-Net enhancement trained on synthetically degraded healthy tissue—to boost anatomical plausibility and fidelity. Empirical results show that simple ensembling yields improvements in $MSE$, $PSNR$, and $SSIM$, while classical post-processing maintains these gains and the approach remains computationally efficient for resource-constrained settings. The proposed pipeline offers a practical path toward robust, accessible clinical deployment and broader adoption of inpainting-based preprocessing in neuroimaging analyses, as demonstrated on the BraTS 2025 inpainting benchmark and Docker deployment notes.

Abstract

Magnetic Resonance Imaging (MRI) is the primary imaging modality used in the diagnosis, assessment, and treatment planning for brain pathologies. However, most automated MRI analysis tools, such as segmentation and registration pipelines, are optimized for healthy anatomies and often fail when confronted with large lesions such as tumors. To overcome this, image inpainting techniques aim to locally synthesize healthy brain tissues in tumor regions, enabling the reliable application of general-purpose tools. In this work, we systematically evaluate state-of-the-art inpainting models and observe a saturation in their standalone performance. In response, we introduce a methodology combining model ensembling with efficient post-processing strategies such as median filtering, histogram matching, and pixel averaging. Further anatomical refinement is achieved via a lightweight U-Net enhancement stage. Comprehensive evaluation demonstrates that our proposed pipeline improves the anatomical plausibility and visual fidelity of inpainted regions, yielding higher accuracy and more robust outcomes than individual baseline models. By combining established models with targeted post-processing, we achieve improved and more accessible inpainting outcomes, supporting broader clinical deployment and sustainable, resource-conscious research. Our 2025 BraTS inpainting docker is available at https://hub.docker.com/layers/aparida12/brats2025/inpt.

Figures (2)

• Figure 1: Overview of the proposed post-processing pipeline for MRI inpainting. Inpainted outputs from the top two models from the BraTS 2024 inpainting challenge are first aggregated using classical tools like pixel-level averaging. The ensemble result is then refined by post-processing, a basic U-Net trained on synthetically blurred healthy brain regions, or a mix of both. This multi-stage pipeline improves anatomical fidelity and reconstruction sharpness while maintaining computational efficiency.
• Figure 2: Qualitative Results Comparison of inpainted brain MR images from different methods visualized in ITK-SNAP. From left to right: Rank 1 (Zhang et al.), Rank 2 (Ferreira et al.), median ensemble, and geometric mean ensemble. All views are centered at the same voxel coordinate (164, 110, 78), but demonstrate differing image intensity values under the cursor: 931.1 (Rank 1), 664.2 (Rank 2), 797.6 (Geometric), and 796.4 (Geometric with median filter postprocessing). Notably, the bottom-left ventricle appears more faded and ill-defined in the Rank 1 output, while it is clearly delineated in the Rank 2 image. The ventricle is slightly more apparent in both the median and geometric ensemble outputs than in Rank 1, indicating partial recovery of anatomical detail through pixel-wise fusion strategies. These variations highlight differences in tissue reconstruction fidelity between individual models and ensemble-based approaches.