Deep Learning Network Embedding Layer for Medical Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Medical image classification in deep learning networks faces challenges due to variability in imaging conditions and the need for image pre-normalization, which can be inconsistent and require manual adjustment, leading to intra- and inter-observer variation and difficulties in handling meta data effectively.
Innovation Solution
An image classification learning network that embeds meta data with image data using a learnable embedding layer, allowing the network to learn correlations between raw image and meta data during training, thereby eliminating the need for pre-normalization and enabling autonomous normalization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If image pre-normalization is performed manually, then image consistency across different imaging conditions is improved, but the process becomes complex and requires manual adjustment leading to intra- and inter-observer variation
Solution Approach 1:
The deep learning network performs normalization automatically during training by learning the correlation between raw image data and meta data, eliminating the need for manual pre-normalization processes. The network self-adjusts to handle imaging condition variations without human intervention.
Solution Approach 2:
The manual mechanical process of image pre-normalization is replaced by an automated computational process where the embedding layer encodes meta data and the network learns relationships between raw images and their corresponding meta data, substituting manual operations with algorithmic processing.
2Reliability
If manual image pre-normalization is applied, then classification consistency is improved, but the process is time-consuming and costly
Solution Approach 1:
The system eliminates time-consuming manual pre-processing by enabling the network to automatically learn and apply normalization transformations during training, significantly reducing the time required for preparation while maintaining classification consistency.
Solution Approach 2:
The network performs preliminary learning of normalization relationships during the training phase on labeled data, so that when new images are processed, the network can immediately apply the learned transformations without requiring time-consuming manual pre-processing steps.
3Measurement precision
If meta data is manually processed and combined with image data, then classification accuracy is improved, but the process becomes complex and difficult to automate
Solution Approach 1:
Manual meta data processing is replaced by an automated embedding layer that encodes meta data into a compact representation and automatically combines it with image data through matrix multiplication operations, simplifying the process while maintaining accuracy.
Solution Approach 2:
The embedding layer merges meta data processing with image data processing by combining the encoded meta data representations with image pixel data through concatenation and matrix multiplication, creating a unified feature space that simplifies subsequent classification operations.
Data Source
AI summary
The present invention relates to a system and method to embed meta data from an imaging and communications system whereby the meta data is combined with image data as an input to a deep learning network. An image classification learning network is disclosed which comprises: a means to input image data and meta data; and an embedding layer comprising learnable embedding weights to encode the meta data to provide a learned object, and a softmax layer to classify a combination of the image data and the learned object.