Improving Robustness and Reliability in Medical Image Classification With Latent-Guided Diffusion and...

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Deep learning models for medical imaging struggle with real-world data corruptions and noise.
• Unreliable predictions and poor confidence calibration hinder clinical adoption of AI in medicine.
• Existing methods address specific issues like robustness or calibration but lack a unified framework.

Purpose of the Study:

• To develop a novel ensemble learning framework, LaDiNE, for robust and reliable medical image classification.
• To improve prediction accuracy and confidence calibration under various image perturbations.
• To bridge the gap between deep learning performance on clean data and real-world clinical applicability.

Main Methods:

• LaDiNE employs an ensemble learning approach combining Vision Transformers and diffusion-based generative models.
• Vision Transformer encoder blocks extract hierarchical, invariant features robust to perturbations.
• Diffusion models act as density estimators for improved data distribution modeling and confidence calibration.

Main Results:

• LaDiNE demonstrated superior performance over state-of-the-art methods on tuberculosis and melanoma datasets.
• The method simultaneously improved prediction accuracy and confidence calibration.
• Effectiveness was shown under unseen noise, adversarial perturbations, and resolution degradation.

Conclusions:

• LaDiNE offers a robust and reliable solution for medical image classification in the presence of data corruptions.
• The combined approach of Vision Transformers and diffusion models enhances model generalizability and trustworthiness.
• This framework has the potential to significantly advance the clinical applicability of AI in medical diagnostics.