Tensorized Network Layers for Medical Image Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional deep network models for medical image classification lose internal structure information and correlation information between image regions, leading to inaccurate classification results.

Innovation Solution

A smart diagnosis assistance method using a triple generative adversarial nets (GAN) model, which includes a tensorized network layer and a second-order pooling module, to preserve and extract discriminative features from medical images.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If a traditional deep network model is used to process medical images, then the processing speed is improved, but the internal structure information and correlation information between image regions are lost, leading to inaccurate classification results

Engineering Contradiction:
Improveprocessing speedVSAvoidinternal structure information and correlation information
Core Design Contradiction:
ProductivityVSLoss of information

Solution Approach 1:

The patent introduces a third dimension (channel dimension) to the pooling operation, transforming traditional 2D pooling into 3D tensor pooling. This allows simultaneous preservation of spatial information (H×W) and channel information (C), resolving the contradiction by adding a dimensional perspective that captures correlations across all three dimensions without sacrificing processing efficiency

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The patent combines multiple types of pooling operations (spatial pooling and channel pooling) into a composite tensor pooling module. This composite approach integrates the advantages of both spatial dimension reduction and channel dimension reduction, preserving both spatial structure and feature correlations while maintaining computational efficiency

Inventive Principle:
Principle #40Composite materials

2Device complexity

If traditional pooling operations are used in the classification model, then the computational complexity is reduced, but the discriminative features and correlation information between channels are lost

Engineering Contradiction:
Improvecomputational complexityVSAvoiddiscriminative features and channel correlation information
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent segments the pooling operation into two distinct stages: spatial pooling (reducing H×W dimensions) and channel pooling (reducing C dimension). This segmentation allows each stage to focus on specific types of information preservation, with spatial pooling maintaining feature maps and channel pooling extracting correlations, thereby reducing overall complexity while preserving essential information

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent extends traditional pooling from 2D spatial operations to 3D tensor operations by incorporating the channel dimension. This dimensional extension enables the model to capture correlations across spatial and channel dimensions simultaneously, preserving discriminative features while maintaining computational tractability through structured reduction

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Measurement precision

If the model uses more parameters to capture internal structure information, then the classification accuracy is improved, but the overfitting problem worsens

Engineering Contradiction:
Improveclassification accuracyVSAvoidrobustness against overfitting
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent changes the parameter representation from traditional flattened vectors to structured tensors that preserve spatial and channel relationships. This parameter transformation allows the model to capture complex patterns with fewer parameters by leveraging the structured nature of medical image data, improving accuracy while reducing overfitting through more efficient parameter utilization

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4064124B1Smart diagnosis assistance method and terminal based on medical images
Publication Date: 2025.05.07 SHENZHEN INST OF ADVANCED TECH CHINESE ACAD OF SCI
  • EP4064124B1 patent drawingFigure 1
  • EP4064124B1 patent drawingFigure 2
  • EP4064124B1 patent drawingFigure 3

AI summary

The present application is suitable for use in the technical field of computers, and provides a smart diagnosis assistance method and terminal based on medical images, comprising: acquiring a medical image to be classified; pre-processing the medical image to be classified to obtain a pre-processed image; and inputting the preprocessed image into a trained classification model for classification processing to obtain a classification type corresponding to the pre-processed image, the classification model comprising tensorized network layers and a second-order pooling module. As the trained classification model comprises tensor decomposed network layers and a second-order pooling module, when processing images on the basis of the classification model, more discriminative features related to pathologies can be extracted, increasing the accuracy of medical image classification.