PCIM: Learning pixel attributions via pixel-wise channel isolation mixing in high content imaging.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Biomedical Imaging

Background:

• Deep Neural Networks (DNNs) excel in computer vision but their black-box nature hinders interpretability, especially in biomedical fields.
• Existing methods for explaining DNN decisions often require access to internal network states or gradients, limiting their applicability.
• There is a significant need for methods that provide clear, pixel-level explanations for DNN decisions to foster trust and adoption.

Purpose of the Study:

• To introduce a novel method, Pixel-wise Channel Isolation Mixing (PCIM), for generating pixel attribution maps.
• To enable the interpretation of DNN decisions without accessing internal network states or gradients.
• To provide a network-agnostic approach for creating interpretable DNNs.

Main Methods:

• PCIM treats each pixel as a distinct input channel.
• A blending layer is trained to mix these pixel channels, reflecting specific classification decisions.
• The method generates pixel attribution maps for individual images, independent of the classification network.

Main Results:

• Benchmark testing on three diverse high-content imaging datasets demonstrated state-of-the-art performance.
• PCIM showed superior model fidelity and localization ability in both fluorescence and bright-field imaging.
• The method successfully generated pixel-level attribution maps for arbitrary DNNs.

Conclusions:

• PCIM offers a unique and effective solution for creating pixel-level attribution maps from any DNN.
• This method enhances the interpretability and trustworthiness of DNNs, particularly in sensitive applications like biomedical imaging.
• PCIM addresses the critical need for explainable AI in computer vision and high-content imaging.