SHFormer: Dynamic Spectral Filtering Convolutional Neural Network and High-pass Kernel Generation Transformer for Adaptive MRI Reconstruction

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Health & Medical Research · Depth: Expert, extended

Summary

SHFormer is a novel deep neural network architecture designed for adaptive MRI reconstruction, addressing limitations of existing attention mechanism (AM)-based models in capturing high-frequency details and generalizing across multimodal data. It integrates a spectral filtering convolutional neural network (SFCNN) for mode-specific transferable features and a dynamic high-pass kernel generation transformer (HKTFB) to focus on high-frequency information. Evaluated across supervised, self-supervised, and diffusion model-based learning, SHFormer demonstrates scalable, high-quality reconstruction. It achieves best improvement margins of ~1 dB in PSNR and ~0.01 in SSIM under unseen scenarios, including closed-set generalization to varying undersampling levels and open-set generalization to unseen MRI contrasts like cardiac, knee, and multi-contrast brain MRI. The model's code is available at https://github.com/sriprabhar/SHFormer.

Key takeaway

For MRI researchers and machine learning engineers developing robust reconstruction models, SHFormer offers a significant advancement. Its ability to propagate high-frequency details and generalize across diverse MRI modalities and undersampling factors means you can achieve higher fidelity reconstructions with reduced retraining needs. Consider integrating its spectral filtering and dynamic high-pass kernel generation principles to enhance your adaptive learning architectures for clinical deployment.

Key insights

SHFormer uses neuromodulation-based spectral and high-pass filtering attention for scalable, high-quality MRI reconstruction across diverse data.

Principles

• Attention mechanisms can struggle with high-frequency details in MRI.
• Neuromodulation enables adaptive learning for heterogeneous MRI data.
• Combining spectral and spatial attention improves generalization.

Method

SHFormer combines a spectral filtering CNN for mode-specific transferable features and a dynamic high-pass kernel generation transformer to propagate context-aware high-frequency details. It uses instance-specific dynamic weight prediction.

In practice

• Apply to supervised, self-supervised, and diffusion MRI reconstruction.
• Generalize models to unseen MRI contrasts and undersampling factors.
• Improve recovery of fine details in cardiac and knee MRI.

Topics

• MRI Reconstruction
• Attention Mechanism
• Spectral Filtering
• High-pass Filtering
• Convolutional Neural Networks
• Transformers
• Multimodal MRI

Code references

• sriprabhar/SHFormer

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• SHFormer: Dynamic Spectral Filtering Convolutional Neural Network and High-pass Kernel Generation Transformer for Adaptive MRI Reconstruction — SHFormer enhances MRI reconstruction by dynamically filtering spectral data and generating high-pass kernels for superior high-frequency detail.
• Revisiting Neural Processes via Fourier Transform and Volterra Series — Combining Volterra series with frequency-domain convolutions yields scalable, interpretable translation-equivariant Neural Processes for irregular data.
• IQ-JEPA: A Joint-Embedding Predictive Architecture with a Hermitian Vision Transformer for Sound Speed and Attenuation Estimation from Ultrasound IQ Data — IQ-JEPA uses a Hermitian Vision Transformer and self-supervision to efficiently estimate ultrasound sound speed and attenuation from IQ data.
• A Multi-Fidelity Convolutional Autoencoder-Transfer Learning Framework for Guided-Wave-Based Damage Diagnosis Using Large Simulated and Limited Experimental Datasets — Multi-fidelity simulation and CAE-based transfer learning overcome data scarcity in GWSHM, achieving high accuracy with limited experimental data.
• Implicit Neural Representations: A Signal Processing Perspective — INRs represent signals as continuous neural network functions, enabling analytical operations and adaptive data modeling.
• MSA-DCNN: A Data-Efficient Multi-Scale Deformable CNN for Medical Image Classification — MSA-DCNN unifies adaptive sampling, multi-scale attention, and self-distillation for data-efficient medical image classification.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.