Abstract
<title>Abstract</title> <p>Volume reconstruction for three-dimensional medical imaging is crucial for clinical analysis but poses serious problems of ill posedness as well as anatomical consistency. State-of-the-art algorithms are inherently limited by their shortcomings: small receptive fields of CNNs, quadratic complexity O(N2) of transformers, and anatomical inconsistency of diffusion models with no accurate uncertainty estimation. We introduce MedMamba3D, a novel hybrid approach that incorporates four complementary techniques into a single architecture with the guidance of uncertainties: (i) a 3D selective state space encoder (Mamba) that has linear complexity O(N) with bidirectional scan; (ii) a deterministic 3D U-Net branch that provides anatomical priors; (iii) a conditional diffusion model for generating high-frequency features; and (iv) a fusion module for merging predictions with optimal inverse variance weighting. Our extensive experiments on BraTS 2021, fastMRI, and MIMIC-CXR show state-of-the-art results (PSNR gain of 3.7 dB against transformers and 3.4 dB against diffusion model), top calibration metrics (ECE=0.08), robust reconstruction for 90% missing data, and clinically significant segmentation improvement (Dice+8.9%).</p>