Neural Steered Mixture of Experts for Medical Image Denoising and Super-Resolution

(Submitted to Pattern Recognition, 2025)
Aytaç Özkan, Thomas Sikora
📄 Access the Paper
⭐ If this work supports your research, please star the repository.

---

Overview

Medical image acquisition is often constrained by hardware limitations and prolonged scanning durations, leading to reduced spatial fidelity. Super-resolution (SR) techniques address these limitations by reconstructing high-resolution (HR) representations from low-resolution (LR) observations, thereby enhancing diagnostic interpretability.

However, conventional single-image super-resolution (SISR) methods often rely on simplified noise assumptions, such as additive white Gaussian noise (AWGN), which fail to capture the non-stationary and modality-dependent degradations present in clinical settings.

We introduce the Neural Steered Mixture of Experts (N-SMoE) framework, a generative adversarial model for joint image denoising and SR. It incorporates:

• A Laplacian resizer and bandpass filtering in the encoder for capturing local and high-frequency structures.
• A probabilistic steered mixture of experts decoder with edge-aware gating mechanisms and autoregressive modeling using 2D adaptive kernels.
• A stochastic degradation model (SDM) to simulate diverse and realistic noise patterns during training.

This architecture yields state-of-the-art (SOTA) performance across multiple medical imaging benchmarks, while maintaining interpretability and robustness across modalities.

---

Features

• Unified framework for denoising and super-resolution in clinical images
• Edge-aware mixture of experts decoder with autoregressive signal modeling
• Stochastic degradation model for training under modality-specific noise
• Generalization across synthetic and real-world datasets

---

Requirements

• Python 3.8
• PyTorch 1.13.0
• Complete environment setup will be provided upon release

---

Code & Models

The training and inference scripts, along with pretrained weights, will be released soon.

---

Citation

Oezkan, Aytac and Sikora, Thomas, Neural Steered Mixture of Experts for Medical Image Denoising, and Super-Resolution. SSRN: https://ssrn.com/abstract=5193694

---

Contact

For questions, contact: [email protected]