Applications of interpretable deep learning in neuroimaging: A comprehensive review

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Machine Learning
• Explainable AI

Background:

• Clinical adoption of deep learning (DL) models is limited by their "black-box" nature, raising concerns about trustworthiness and reliability.
• These concerns are amplified in neuroimaging due to complex brain phenotypes and inter-subject variability.
• Interpretable deep learning (iDL) offers a solution by visualizing and explaining DL model operations.

Purpose of the Study:

• To systematically review neuroimaging applications of iDL methods.
• To critically analyze the evaluation of iDL explanation properties in existing literature.
• To identify optimal iDL approaches for neuroimaging data and suggest future research directions.

Main Methods:

• Conducted a systematic literature review of neuroimaging studies employing iDL methods.
• Identified and categorized ten distinct types of iDL methods used in the reviewed studies.
• Analyzed five key properties of iDL explanations: biological validity, robustness, continuity, selectivity, and downstream task performance.

Main Results:

• Seventy-five studies were included in the review.
• Ten categories of iDL methods were identified.
• The most prevalent iDL methods in current literature may be suboptimal for neuroimaging data analysis.

Conclusions:

• Despite the potential of iDL to enhance trust in AI for neuroimaging, current popular methods require re-evaluation.
• Further research is needed to develop and validate iDL approaches specifically tailored for the complexities of neuroimaging data.
• Optimizing iDL methods is crucial for advancing the reliable clinical application of AI in neuroscience.