AI-based association analysis for medical imaging using latent-space geometric confounder correction

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computational Biology

Background:

• Confounding effects pose challenges in AI-based medical image analysis, potentially leading to misinterpretations.
• Existing methods for confounder removal can degrade image reconstruction quality, limiting feature visualization.
• Interpretability and accurate association analysis are crucial for clinical and epidemiological research.

Purpose of the Study:

• To develop a novel AI strategy for medical image analysis that effectively addresses confounding effects.
• To propose a method that retains essential information while generating confounder-free latent representations.
• To enhance the interpretability of AI models in medical imaging for reliable feature visualization.

Main Methods:

• A novel approach utilizing autoencoder latent space as a vector space is proposed.
• A correlation-based loss function is introduced to find confounder-free vectors orthogonal to confounder vectors.
• The method encourages linear correlation between latent representations and imaging variables.

Main Results:

• Demonstrated efficacy across three diverse applications with multiple confounders and image modalities.
• Successfully reduced confounder influences and prevented misleading associations in medical image analysis.
• Provided unique visual interpretations of confounder-free representations for in-depth research.

Conclusions:

• The proposed method offers an effective solution for confounding in AI-based medical image analysis.
• It enhances model interpretability and supports reliable feature visualization without compromising image quality.
• This approach facilitates more accurate investigations for clinical and epidemiological researchers.