Plenoptima

In medical image analysis (MI), achieving high-resolution detail is crucial for accurate diagnoses, but remains restricted by extended acquisition times and the need for high frame rates. Existing deep learning-based superresolution (SR) techniques, which often employ simplified degradation models (e.g., Additive White Gaussian Noise (AWGN)), fail to capture the complex, modality-dependent noise characteristics encountered in clinical settings. To address these limitations, we introduce a Neural Parametric Steered Mixture of Experts (N-SMoE) framework, trained within a Generative Adversarial Network (GAN) paradigm, and incorporating an advanced Stochastic Degradation Model (SDM). The SDM applies diverse levels of blurring and noise to downsampled images, thus simulating realistic clinical conditions and mitigating training instability. In contrast to conventional bilinear or bicubic interpolation approaches, the N-SMoE encoder leverages a lightweight, multilayer Laplacian resizer with bandpass filtering capabilities, while a multi-head attention module isolates high-frequency structural patterns to estimate Gaussian primitives for the decoder. The SMoE decoder, unlike conventional deep learning (DL) black-box models, utilizes gating functions and two-dimensional kernels to produce an edge-aware representation capable of modeling both continuous and smooth transitions, which makes it an excellent tool for SR-related restoration tasks. The resulting proposed neural parametric autoregressive framework enhances interpretability and surpasses state-of-the-art (SotA) methods in terms of fidelity, perceptual quality, and subjective evaluations across multiple medical imaging modalities.