Capsule Endoscopy Image Classification via Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing capsule endoscopy systems generate a large number of images, mostly redundant or lacking details, which burdens physicians with the task of identifying specific anatomical structures within the digestive tract, necessitating an efficient method to classify and filter these images.
Innovation Solution
A transfer learning-based capsule endoscopic images classification system comprising an image preprocessing module, similar image removal module, and key image positioning module, which removes redundant images using brightness thresholds, optical flow for similar image detection, and combines deep learning features with manual extraction to classify images by anatomical structure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If capsule endoscope takes a large number of images to ensure coverage of digestive tract, then the completeness of anatomical structure detection is improved, but the workload of physicians and time required for review increases significantly
Solution Approach 1:
The patent extracts and removes redundant images from the large set of captured images through multiple filtering stages: brightness threshold filtering removes too bright or dark images, detail detection removes blurry images, and similarity comparison removes duplicate images. This extraction of useful information from the large dataset reduces the number of images physicians need to review while maintaining complete anatomical structure detection.
Solution Approach 2:
The system performs preliminary processing of images before physician review by automatically classifying them into anatomical structures (esophagus, stomach, duodenum, jejunum, ileum, cecum, colon, rectum) and organizing them in chronological order. This preliminary classification action prepares the data in advance, allowing physicians to quickly locate specific anatomical regions without manually searching through all images.
2Adaptability or versatility
If capsule endoscope captures all regions of digestive tract, then the coverage of anatomical structures is improved, but the quantity of images to be processed increases
Solution Approach 1:
The patent changes parameters of images for comparison by adjusting brightness to a standard range and extracting feature vectors that represent key characteristics. This parameter transformation allows efficient comparison and classification of images across different anatomical structures while reducing the effective data volume that needs to be processed and stored.
Solution Approach 2:
The system segments the large set of images into distinct anatomical structure categories (esophagus, stomach, duodenum, jejunum, ileum, cecum, colon, rectum) using deep learning classification. This segmentation organizes the large quantity of images into manageable groups, making it easier to process and review specific regions while maintaining comprehensive coverage of the entire digestive tract.
3Measurement precision
If physicians review all captured images to identify specific anatomical structures, then the accuracy of anatomical structure identification is improved, but the efficiency of work decreases
Solution Approach 1:
The patent introduces an intermediary automated classification system that uses deep learning models to analyze images and identify anatomical structures. This intermediary system processes images with high accuracy and presents results to physicians in an organized manner, maintaining identification accuracy while dramatically improving work efficiency by eliminating the need for physicians to manually examine every image.
4Reliability
If similar images are kept to ensure no diagnostic information is lost, then the completeness of diagnostic data is improved, but the redundancy of data increases
Solution Approach 1:
The system uses feedback from similarity comparison results to determine whether to keep or remove images. By comparing feature vectors of consecutive images and calculating similarity metrics, the system provides feedback on redundancy levels and makes intelligent decisions to remove only truly duplicate images while preserving images that contain unique diagnostic information, thus maintaining data completeness while reducing redundancy.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This system significantly reduces the number of images for physician review by removing too bright or dark images, images without details, and similar images, thereby improving classification efficiency and facilitating the identification of key anatomical structures.
Implementation Method 1
The similar image removal module removes similar images from the capsule endoscopic images using optical flow method
Data Source
AI summary
The present invention provides a transfer learning based capsule endoscopic images classification system. The system removes the capsule endoscopic images with an average brightness value beyond the preset threshold, and removes the capsule endoscopic images without details based on image brightness standard deviation and image brightness gradient. The system also removes similar images from the capsule endoscopic images using optical flow method, classifies the capsule endoscopic images according to the corresponding anatomical structure, and obtains the classified capsule endoscopic images list arranged in chronological order. The system further determines and labels the position of the first image of each specific anatomical structure in the classified capsule endoscopic images list arranged in chronological order.