Capsule Endoscope Image Recognition Using 3D CNN Video Sequences

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing capsule endoscope image recognition methods struggle with low accuracy due to the inability to utilize spatial and temporal information from multiple images simultaneously, leading to confusion between gastrointestinal folds and lesions like polyps and bulges, especially when captured at specific angles.

Innovation Solution

A capsule endoscope image recognition method using a 3D convolutional neural network that processes multiple frames of images, segmenting them into groups using a sliding window technique, and combines RGB and optical flow data to improve recognition accuracy through a collaborative spatio-temporal feature structure.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If single image recognition is used, then the recognition process is simple, but the recognition accuracy is low due to inability to utilize spatial and temporal information from multiple images

Engineering Contradiction:
Improvelesion recognition accuracyVSAvoidrecognition system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent transitions from 2D single-image recognition to 3D video sequence recognition by adding the temporal dimension. The 3D CNN model processes video clips consisting of multiple frames, enabling simultaneous extraction of spatial features (from individual frames) and temporal features (from frame sequences), thereby improving lesion recognition accuracy through multi-dimensional feature analysis

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

Solution Approach 2:

The patent divides the video sequence into multiple frames and processes them through segmented computational pathways. The sliding window technique segments video sequences into overlapping clips, and the dual-branch CNN architecture segments feature extraction into spatial and temporal components, allowing comprehensive analysis while managing computational complexity

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If multiple images are processed to obtain spatial and temporal information, then the recognition accuracy improves, but the computational complexity and processing time increase

Engineering Contradiction:
Improvelesion identification accuracyVSAvoidimage processing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary processing by pre-extracting spatial features from individual frames using 2D convolutional layers before temporal feature extraction. This preliminary spatial feature extraction reduces the dimensionality of subsequent temporal processing, optimizing the balance between comprehensive analysis and processing efficiency

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent employs dynamic feature extraction where the importance of spatial versus temporal features is adaptively determined. The model dynamically adjusts feature weighting based on the specific lesion type and video sequence characteristics, optimizing processing efficiency for different diagnostic scenarios

Inventive Principle:
Principle #15Dynamics

Data Source

PatentEP4276684B1Capsule endoscope image recognition method based on deep learning, and device and medium
Publication Date: 2026.04.22 ANKON TECHNOLOGIES CO LTD
  • EP4276684B1 patent drawingFigure 1
  • EP4276684B1 patent drawingFigure 2
  • EP4276684B1 patent drawingFigure 3

AI summary

The present invention discloses a capsule endoscope image recognition method based on deep learning, and a device and a medium. After a plurality of frames of continuously captured images are processed to form an image sequence of a specific format, multi-channel recognition is performed on the plurality of frames of images by means of a 3D convolutional neural network model, and a recognition result is then output in combination with a recognition probability of each channel, such that the image recognition precision is improved.