Signature Activation: A Sparse Signal View for Holistic Saliency

TL;DR

This paper introduces Signature Activation, a saliency method for CNNs that provides holistic, class-agnostic explanations, particularly effective for medical images like angiograms, with potential clinical applications.

Contribution

It proposes a novel saliency technique based on a sparse signal view, with theoretical justification and demonstrated clinical relevance for lesion detection.

Findings

Effective in highlighting lesions in coronary angiograms

Provides holistic, class-agnostic explanations

Theoretically justified approach

Abstract

The adoption of machine learning in healthcare calls for model transparency and explainability. In this work, we introduce Signature Activation, a saliency method that generates holistic and class-agnostic explanations for Convolutional Neural Network (CNN) outputs. Our method exploits the fact that certain kinds of medical images, such as angiograms, have clear foreground and background objects. We give theoretical explanation to justify our methods. We show the potential use of our method in clinical settings through evaluating its efficacy for aiding the detection of lesions in coronary angiograms.