X-Ray Super-Resolution Reconstruction With Hallucination Filtering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

X-ray Microscopy Imaging faces challenges in achieving high-resolution imaging of large samples due to noise and artifacts, with deep-learning-based techniques introducing hallucinations that complicate objective resolution assessment.

Innovation Solution

The method involves calculating a relative modulation transfer function (RMTF) to filter out hallucinated frequency components, using a trained neural network to improve volumetric reconstructions, and applying a masking function to enhance image quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If deep-learning-based image processing techniques are used to improve resolution and reduce noise, then image quality is improved, but hallucinations are introduced that complicate objective resolution assessment

Engineering Contradiction:
Improveimage qualityVSAvoidobjective resolution assessment
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent introduces a baseline volumetric representation as an intermediary reference to objectively assess the neural network's output. By comparing the improved reconstruction against this baseline, the system can distinguish genuine resolution improvements from hallucinations, restoring objective measurement capability while maintaining the quality benefits of deep learning

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements a feedback mechanism where the neural network's output is evaluated against a baseline representation, and this evaluation information is used to guide further processing. The feedback loop enables continuous refinement and objective assessment of resolution improvement, preventing unchecked hallucination generation

Inventive Principle:
Principle #23Feedback

2Measurement precision

If high-resolution tomography acquisition is used to achieve fundamental structure imaging, then resolution is improved, but acquisition time increases significantly

Engineering Contradiction:
ImproveresolutionVSAvoidacquisition time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary action by training the neural network offline using paired high-resolution and low-resolution data. This pre-training phase captures the resolution enhancement patterns, allowing the network to subsequently process low-resolution acquisitions and generate high-resolution outputs without requiring time-consuming high-resolution scanning

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses low-resolution acquisitions as proxies or copies that can be processed by the neural network to generate high-resolution results. Instead of directly acquiring time-consuming high-resolution data, the system creates a computational copy through the trained network that replicates the desired high-resolution output

Inventive Principle:
Principle #26Copying

3Area of stationary object

If low-resolution detectors are used to achieve large field of view, then field of view is improved, but sensitivity to high-energy X-rays deteriorates

Engineering Contradiction:
Improvefield of viewVSAvoiddetector sensitivity
Core Design Contradiction:
Area of stationary objectVSReliability

Solution Approach 1:

The patent replaces the mechanical/detector-based resolution limitation with a computational solution. Instead of relying on detector physics to provide high resolution, the system uses a neural network to computationally enhance resolution from low-resolution detector data, substituting algorithmic processing for physical detector limitations

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP4657380A1Robust multiscale x-ray super-resolution reconstruction
Publication Date: 2025.12.03 CARL ZEISS X-RAY MICROSCOPY INC
  • EP4657380A1 patent drawingFigure 1
  • EP4657380A1 patent drawingFigure 2
  • EP4657380A1 patent drawingFigure 3

AI summary

A technique is disclosed for analyzing and displaying the extent to which the images and structures inferred by a physically seeded multiscale network correspond to genuine resolution improvement through noise insensitive point spread function deconvolution, and the extent to which they correspond to the hallucination of realistic looking structures with realistic frequency contents. A relative modulation transfer function can be computed, which can represent the distribution of frequency components in a particular reconstruction (e.g., a volumetric reconstruction from high-resolution data) that are not robustly recovered by a different reconstruction (e.g., a volumetric reconstruction via processing of low-resolution data with a trained neural network). The high-frequency portion of these frequency components can represent hallucinations introduced by a trained neural network, and can be leveraged to filter the different reconstruction prior to further use.