Mass Image Quality Conversion Using Cross-Modality Transfer Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The challenge in improving the image quality of mass images is exacerbated by the small amount of ions acquired from each micro area, leading to low signal-to-noise ratios and difficulty in applying machine learning-based image quality conversion models, as high-definition mass images are hard to obtain.

Innovation Solution

A mass image processing apparatus comprising a pre-processor, converter, and post-processor that applies intensity scaling and noise correction processes to fit the image quality conversion model's input conditions, allowing for effective transfer learning from scanning electron microscope images to enhance mass image quality, even when high-definition mass images are not available.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Object-affected harmful factors

If a simple smoothing filter is applied to reduce noise in mass images, then noise prominence is reduced, but the entire mass image becomes blurred

Engineering Contradiction:
Improvenoise prominenceVSAvoidimage clarity
Core Design Contradiction:
Object-affected harmful factorsVSMeasurement precision

Solution Approach 1:

The patent introduces a machine learning model as an intermediary between the noisy mass image and the final processed image. This model, trained on paired low-definition and high-definition images from other modalities, acts as a mediator that can suppress noise while preserving or enhancing fine details, avoiding the blurring effect of traditional smoothing filters.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent changes the fundamental parameter of image processing from simple spatial filtering to complex machine learning-based transformation. By training a neural network model to learn the mapping between low-definition and high-definition images, the system transforms the processing approach from local pixel manipulation to global pattern recognition, enabling noise reduction without detail loss.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If machine learning-based image quality conversion models are used to improve mass image quality, then image quality can be enhanced, but high-definition mass images are difficult to obtain for model training

Engineering Contradiction:
Improveimage qualityVSAvoidavailability of training data
Core Design Contradiction:
Measurement precisionVSEase of manufacture

Solution Approach 1:

The patent uses images from other modalities (such as optical microscopy or electron microscopy) as substitutes or copies for the unavailable high-definition mass images. These alternative images serve as proxies to train the machine learning model, allowing the system to learn the transformation to high-definition images without requiring actual high-definition mass spectrometry images for training.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent applies a universal machine learning model that can be trained on data from multiple sources and modalities. The model learns general image enhancement patterns that are transferable to mass images, making the training process more flexible and reducing dependency on specific high-definition mass image datasets.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If high-definition mass images are acquired to train image quality conversion models, then model training can be performed, but the small amount of ions from each micro area makes this extremely difficult

Engineering Contradiction:
Improveimage definitionVSAvoidion amount
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent uses images from other imaging modalities as substitutes for the unavailable high-definition mass images. These alternative images provide the necessary training data without requiring additional ion acquisition, effectively copying the structural information from one modality to train the model for another modality.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent introduces a machine learning model trained on alternative modality images as an intermediary that bridges the gap between low ion amount measurements and high-definition image requirements. This intermediary translates information from modalities with sufficient signal to the mass imaging domain.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4283656B1Mass image processing apparatus, method and program
Publication Date: 2024.09.25 JEOL LTD
  • EP4283656B1 patent drawingFigure 1
  • EP4283656B1 patent drawingFigure 2
  • EP4283656B1 patent drawingFigure 3

AI summary

A pre-processor (46) applies a pre-process to an original mass image produced through mass spectrometry of a sample, to produce a model input image. An image quality converter (48) has an image quality conversion model produced through machine learning based on a group of images produced by a scanning electron microscope, and produces a model output image through image quality conversion of the model input image. A post-processor (50) applies a post-process to the model output image, to produce a mass image after image quality conversion.