The Latent Doctor Model for Modeling Inter-Observer Variability.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Medical imaging tasks often have ambiguous interpretations, leading to inter-observer variability and high variance in reference standards.
• Current methods commonly discard expert uncertainty by training on consensus or majority labels, losing valuable information.
• Developing models that leverage the full distribution of expert labels is crucial for accurate medical image analysis.

Purpose of the Study:

• To develop a novel framework for training on the full label distribution in medical imaging.
• To predict both the uncertainty within expert panels and the most likely ground-truth label.
• To improve the handling of inter-observer variability in medical image analysis.

Main Methods:

• Proposed a new stochastic classification framework: the Latent Doctor Model (LDM).
• LDM is based on a conditional variational auto-encoder architecture.
• Conducted comparative analyses against majority vote models and other label distribution learning methods.

Main Results:

• The LDM significantly outperformed the majority vote baseline in reproducing reference-standard label distributions.
• LDM demonstrated superior performance in modeling label distributions and uncertainty for prostate tumor grading tasks compared to other baselines.
• LDM showed competitive performance against computationally intensive deep ensembles in tumor budding classification.

Conclusions:

• The Latent Doctor Model effectively models expert label distributions and associated uncertainty in medical imaging.
• LDM offers a robust approach to handling inter-observer variability, outperforming conventional methods.
• This framework advances the potential of AI in medical image analysis by incorporating nuanced expert opinions.