Medical Image Embedding Clustering for Data Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accumulation of redundant medical images leads to inefficiencies in data analysis, storage, and machine-learning algorithm training, as most image data captured is of normal states rather than abnormal conditions, resulting in resource wastage and skewed algorithm training.
Innovation Solution
A computing device with an image embeddings generating module, clustering module, and replacement determining module that groups images into clusters, replacing each cluster's images with a representative image, reducing data volume and improving algorithm training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If all captured medical images are stored and analyzed, then complete data availability is achieved, but storage space and processing time are excessively consumed
Solution Approach 1:
The patent extracts only the essential information from medical images by generating image embeddings that capture the semantic meaning of each image. Instead of storing all raw images, the system extracts and stores only the embedding representations, which contain the crucial diagnostic information while occupying minimal storage space.
Solution Approach 2:
The patent transforms images from their original high-dimensional pixel space into a compressed embedding space with reduced dimensions. This parameter transformation maintains the essential diagnostic features while dramatically reducing the data volume, allowing efficient storage and processing without losing critical medical information.
2Loss of information
If all captured medical images are stored and analyzed, then complete data availability is achieved, but processing time and computational resources are excessively consumed
Solution Approach 1:
The system extracts only the essential semantic features from each medical image and stores them as embeddings. During analysis, only these compact embedding representations are processed rather than the full-resolution images, dramatically reducing computational time and resource consumption while preserving diagnostic accuracy.
Solution Approach 2:
The patent performs preliminary processing by generating image embeddings in advance and storing them. This pre-extraction of essential features allows for rapid querying and analysis later, as the system only needs to process the pre-computed embeddings rather than analyzing full images during critical diagnostic operations.
3Quantity of substance
If a large number of identical and redundant images are used for training, then data volume is increased, but machine-learning algorithm accuracy is skewed towards normal states
Solution Approach 1:
The patent extracts the essential semantic content from images using embeddings, which allows the system to identify and filter out redundant duplicate images. By working with embedding representations rather than raw images, the system can efficiently detect and remove duplicates while preserving the diversity of abnormal cases for training, thereby improving algorithm accuracy.
Solution Approach 2:
The transformation to embedding space enables more effective identification and removal of redundant images. The compact representation allows for efficient similarity comparison and duplicate detection, ensuring that the training dataset contains diverse abnormal cases without unnecessary repetitions of normal states, thus improving model reliability.
4Loss of information
If redundant images are transmitted for storage, then data completeness is maintained, but bandwidth consumption is excessively high
Solution Approach 1:
The patent transforms images into compressed embedding representations, dramatically reducing the data size that needs to be transmitted. This parameter transformation from pixel space to embedding space maintains the essential diagnostic information while reducing bandwidth consumption by orders of magnitude, enabling efficient data transfer and storage.
Data Source
AI summary
A computing device includes: an input interface configured to receive a plurality of images of a medical scene; an image embeddings generating module configured to receive, as its input, the plurality of images and to generate a data array as an image embedding for each image; a clustering module configured to determine a plurality of clusters within the plurality of images based on the generated image embeddings; a replacement determining module configured to determine, for each cluster of the determined plurality of clusters, a replacement image; and an output module configured to generate output data wherein, for each determined cluster, all images of that cluster are replaced with the replacement image for that cluster.