Pixel-Precise Explanation Map Generation for Deep Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for ascertaining explanation maps in medical image diagnostics using deep neural networks fail to provide a pixel-precise representation, making it difficult for physicians to identify significant image regions contributing to classification results.
Innovation Solution
A method that generates an explanation map by optimizing a loss function with secondary conditions, ensuring the map's activation does not exceed the original image's activation, using a multiplicative or additive mask to select relevant pixels, and retaining edge and color information, thereby allowing for accurate human interpretation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If explanation maps are generated using conventional deep neural network methods, then classification results are obtained, but pixel-precise representation of significant image regions is not achieved
Solution Approach 1:
The explanation map is segmented into two distinct components: a relevance map indicating significant regions and a mask preserving original pixel values. This segmentation allows the system to identify important areas while retaining precise edge and color information from the original image, resolving the contradiction between precision and information loss.
Solution Approach 2:
The patent introduces a new dimensional structure by creating an explanation map in the same spatial dimensions as the input image rather than in activation map dimensions. This dimensional alignment enables pixel-precise representation while preserving original image characteristics through the mask component.
2Measurement precision
If the explanation map is created in the size of the incoming image, then accurate representation with edge and color information is achieved, but the activation may exceed that of the original image
Solution Approach 1:
The patent implements dynamic control of the explanation map through iterative optimization that adjusts pixel values based on activation constraints. The system dynamically balances location accuracy with activation reliability by adjusting the mask values to ensure the explanation map's activation does not exceed the original image's activation while maintaining precise location information.
Solution Approach 2:
The optimization process incorporates feedback mechanisms where the activation levels of the explanation map are continuously monitored and adjusted. The loss function provides feedback signals that guide the adjustment of mask values, ensuring that location accuracy is maintained while activation levels remain within reliable bounds.
3Loss of information
If all pixels are changed based on significance for classification, then relevant regions are highlighted, but the classification result may change
Solution Approach 1:
The patent applies local quality modification by changing only those pixel values in the explanation map that correspond to regions significant for classification. The mask selectively modifies pixels based on their relevance to the classification task, preserving classification consistency while extracting and highlighting relevant information through localized changes rather than global modification.
Data Source
AI summary
A method for ascertaining an explanation map of an image. All those pixels of the image are highlighted which are significant for a classification of the image ascertained with the aid of a deep neural network. The explanation map is being selected in such a way that it selects a smallest possible subset of the pixels of the image as relevant. The explanation map leads to the same classification result as the image when the explanation map is supplied to the deep neural network for classification. The explanation map is selected in such a way that an activation caused by the explanation map does not essentially exceed an activation caused by the image in feature maps of the deep neural network.