CNN Lesion Detection in Device-Assisted Enteroscopy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine learning techniques for automatic detection and classification of esophagus, stomach, and colon lesions in device-assisted enteroscopy images face challenges in achieving high accuracy and low false-negative rates, which can lead to inappropriate treatment and clinical outcomes.

Innovation Solution

A deep learning-based method using convolutional neural networks (CNNs) with transfer learning and semi-supervised learning approaches, employing architectures like VGG16, InceptionV3, and EfficientNet, to detect and classify lesions such as blood or hematic residues, ulcers, erosions, and vascular lesions, achieving high accuracy and continuous improvement.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If current machine learning techniques are used for automatic detection and classification of lesions, then the system can process images automatically, but the accuracy is insufficient and false-negative rates are high

Engineering Contradiction:
Improvedetection accuracyVSAvoidfalse-negative rate
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The system performs preliminary actions by pre-processing images to enhance lesion visibility and extracting multiple types of features (color, texture, shape) before classification. This preliminary feature extraction and image enhancement prepares the data in optimal form for detection, improving both accuracy and reducing false negatives by ensuring quality input data for the classification algorithms.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system employs composite approaches by combining multiple machine learning techniques (support vector machines, random forests, neural networks) and multiple feature types (color, texture, shape features) into an integrated classification system. This composite methodology leverages the strengths of different algorithms and feature types to achieve higher detection accuracy and lower false-negative rates than any single technique alone.

Inventive Principle:
Principle #40Composite materials

2Measurement precision

If deep learning models with multiple architectures are trained, then classification accuracy improves to 95%, but computational complexity and training time increase

Engineering Contradiction:
Improveclassification accuracyVSAvoidmodel architecture complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments the deep learning process into distinct phases: feature extraction using pre-trained models (VGG16, InceptionV3, EfficientNet), followed by separate classification stages. This segmentation allows each component to be optimized independently - using transfer learning for feature extraction reduces the complexity burden on the final classification models while maintaining high accuracy through the power of multiple architectures.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary action through transfer learning, where pre-trained models on large datasets (ImageNet) provide initial feature extraction capabilities. This preliminary feature extraction by established architectures reduces the training complexity and computational burden on the task-specific models, allowing the system to achieve 95% accuracy without proportionally increasing overall system complexity.

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If transfer learning and semi-supervised learning are employed, then the system achieves high accuracy with limited labeled data, but the training process becomes more complex

Engineering Contradiction:
Improvedetection accuracyVSAvoidtraining process complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system performs preliminary action through transfer learning by initializing models with weights pre-trained on large-scale image datasets. This preliminary training on abundant unlabeled or weakly-labeled data provides strong feature extraction capabilities that can be fine-tuned with limited labeled medical images, achieving high accuracy while the pre-training phase absorbs much of the computational complexity.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system uses semi-supervised learning as an intermediary approach, where a small amount of labeled data guides the training process while leveraging information from large amounts of unlabeled data. This intermediary labeled dataset acts as a bridge, allowing the model to learn from abundant unlabeled images without requiring extensive manual annotation, thus achieving high accuracy with manageable training complexity.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4384979B1Automatic detection and differentiation/classification of the esophagus, stomach, small bowel and colon lesions in device-assisted enteroscopy using a convolutional neuronal network
Publication Date: 2024.12.18 DIGESTAID ARTIFICIAL INTELLIGENCE DEV LDA
  • EP4384979B1 patent drawingFigure 1
  • EP4384979B1 patent drawingFigure 2
  • EP4384979B1 patent drawingFigure 3

AI summary

The present invention relates to a computer-implemented method capable of automatically detecting esophageal, stomach, small bowel and colon lesions, such as blood or hematic residues, small bowel ulcers and erosions, small bowel protruding and vascular lesions in device-assisted enteroscopy image/videos data, by classifying pixels as lesion or non-lesion, using a convolutional image feature extraction step followed by a classification step and indexing such lesions in the set of one or more classes.