3D MRI Inpainting for Structural Reconstruction of Brain Tumor Regions via Composite Spatial–Frequency Optimization

Overview

Breast cancer is the leading cause of cancer-related death among women, with approximately 42,000 deaths annually in the United States and over 700,000 worldwide. A majority of these deaths result from metastatic spread rather than the primary tumor. Brain metastases represent one of the most devastating complications, developing in 20–40% of patients with aggressive subtypes and contributing significantly to mortality in young women aged 20–49. Conventional MRI demonstrates limited sensitivity for sub-centimeter lesions and frequently produces blurred or poorly defined tumor margins. These edge ambiguities critically compromise stereotactic radiosurgery (SRS), where radiation dose is mapped directly to tumor boundaries and even sub-millimeter inaccuracies can result in neurocognitive decline or radiation necrosis. Despite this urgent clinical need, no reconstruction-based framework exists to restore tumor regions from surrounding brain context to improve volumetric edge clarity for radiation planning. This work introduces the first dedicated 3D structural reconstruction framework for breast cancer brain metastases (BCBM). This is the first model to perform full volumetric tumor inpainting in 3D MRI, the first to reconstruct missing tumor regions from masked inputs using contextual brain anatomy, and the first to explicitly optimize tumor edge precision within a reconstruction objective. Unlike prior work limited to denoising or 2D enhancement, this framework restores anatomically plausible tumor structure in three dimensions to support millimeter-level treatment planning. A systematic architectural benchmark was first conducted across six volumetric deep learning models to identify the strongest reconstruction backbone. Among CNN, standard 3D U-Net, SwinUNETR, and other baselines, 3D UNet++ demonstrated the highest baseline fidelity (PSNR 21.96, SSIM 0.5964, MSE 0.0086). Building upon this optimal backbone, we introduced a fully enhanced architecture, called 3D CerebraNet++, incorporating attention-augmented dense skip connections, residual learning, instance normalization with affine adaptation, PReLU activation, SE channel recalibration, CBAM spatial attention, deep supervision, and spatial dropout. Beyond being the first in clinical application, this work is also the first to introduce multiple novel machine learning contributions. It presents the first application of CutMix augmentation in full 3D MRI reconstruction, the first multi-orientation 2.5D VGG perceptual supervision strategy for tumor inpainting, and the first unified composite loss integrating spatial-domain, frequency-domain, perceptual, and edge-aware gradient constraints in a volumetric reconstruction setting. This establishes a new methodological benchmark not only in neuro-oncologic imaging but in 3D deep learning rigor. The pipeline processes NIfTI-based MRI volumes using isotropic 1 mm³ resampling, Z-score normalization, percentile clipping, intensity scaling, and tumor-centered 96×96×96 patch extraction. Optimization employs AdamW with warmup cosine annealing, mixed precision training, gradient clipping, stochastic weight averaging, and early stopping to ensure stability and generalization. Following architectural enhancement and multi-domain supervision, SSIM improved from 0.5964 (baseline 3D UNet++) to 0.98 in 3D CerebraNet++, representing a substantial gain in structural fidelity and boundary clarity. Evaluation was conducted using 5-fold cross-validation with rigorous non-parametric statistical validation, including bootstrap confidence intervals, Kruskal-Wallis fold consistency testing, Wilcoxon signed-rank threshold analysis, Shapiro-Wilk normality assessment, and coefficient of variation analysis. Statistical consistency across folds confirms robustness and reproducibility. This work establishes the first benchmark for volumetric BCBM structural reconstruction, the first tumor-edge-aware 3D reconstruction framework optimized for stereotactic radiosurgery precision, and a new standard of technical rigor in volumetric deep learning. By combining clinical necessity with architectural innovation and statistical validation, this research advances both translational neuro-oncology and the frontier of 3D machine learning methodology.