Visual attribution using Adversarial Latent Transformations.

Area of Science:

• Medical imaging analysis
• Computer vision
• Artificial intelligence in healthcare

Background:

• Accurate localization of disease indicators in medical images is crucial for understanding disease progression and for weakly-supervised segmentation.
• Current methods, relying on class-salient regions or adversarial image-to-image translation, have limitations in capturing all relevant features, are prone to data biases, or generate inefficient solutions.

Purpose of the Study:

• To propose a novel approach, Visual Attribution using Adversarial Latent Transformations (VA2LT), to address the limitations of existing methods for disease indicator localization.
• To accurately locate disease indicators within medical images for improved disease comprehension and diagnostic correlator analysis.

Main Methods:

• VA2LT employs adversarial learning to generate counterfactual (CF) normal images from abnormal images by identifying and altering discrepancies in the latent space.
• Cycle consistency between query and CF latent representations is utilized to guide the training process.

Main Results:

• The proposed VA2LT method was evaluated on three diverse datasets: a synthetic dataset, the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, and the BraTS dataset.
• VA2LT demonstrated superior performance compared to baseline and related methods across all evaluated datasets.

Conclusions:

• VA2LT offers a more effective and robust approach for visual attribution and disease indicator localization in medical imaging.
• The method's ability to generate CF normal images through latent space manipulation shows significant promise for advancing medical image analysis and diagnosis.