Explainable Visual Attention for Deep Learning Medical Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning techniques for medical image analysis lack interpretability and explainability, as attention scores are hidden within the network and carry no readily-understandable physical meaning.
Innovation Solution
A system that includes a deep learning neural network with a primary processing channel for inferencing tasks and an auxiliary processing channel that generates pixel-wise or voxel-wise attention maps, trained using ground-truth attention maps to enhance interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep learning neural network is trained to perform inferencing tasks on medical images, then the network can automatically identify and focus on important portions of the image, but the attention scores generated are hidden within the network and lack interpretability
Solution Approach 1:
The patent extracts the attention mechanism from the hidden internal layers of the neural network and presents it as a separate, interpretable output component. The attention map is generated by an auxiliary processing channel that branches off from the primary channel, allowing the attention scores to be extracted and displayed as visual overlays on the medical images, making them externally observable and interpretable rather than hidden within the network
Solution Approach 2:
The patent introduces an auxiliary processing channel as an intermediary component between the input medical image and the final inferencing output. This auxiliary channel processes the image to generate attention maps that serve as intermediaries, providing human-interpretable information about which regions the network focused on, while the primary channel continues to perform the inferencing task
2Measurement precision
If an auxiliary processing channel is added to generate attention maps, then interpretability is improved, but device complexity increases
Solution Approach 1:
The patent segments the neural network into two distinct processing channels: a primary processing channel that performs the inferencing task and an auxiliary processing channel that generates attention maps. This segmentation allows each channel to have specialized functionality, with the auxiliary channel dedicated solely to producing interpretable attention information, while the primary channel focuses on the inferencing objective
Solution Approach 2:
The auxiliary processing channel serves multiple functions: it processes the input medical image, generates pixel-wise or voxel-wise attention maps, and provides interpretability feedback. By making this component multi-functional, the patent reduces the need for separate dedicated components for each function, thereby managing complexity while improving interpretability
Data Source
AI summary
Systems or techniques that facilitate explainable visual attention for deep learning are provided. In various embodiments, a system can access a medical image generated by a medical imaging scanner. In various aspects, the system can perform, via execution of a deep learning neural network, an inferencing task on the medical image. In various instances, the deep learning neural network can receive as input the medical image and can produce as output both an inferencing task result and an attention map indicating on which pixels or voxels of the medical image the deep learning neural network focused in generating the inferencing task result.