ReaderAdaptNet: modeling reader variability in breast imaging with reader-specific embeddings

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Inter-reader variability is a significant challenge in breast imaging interpretation, affecting tasks like breast density and background parenchymal enhancement (BPE) classification.
• Subjective interpretation leads to inconsistencies, limiting the reliability of AI models trained on aggregated or noisy labels.

Purpose of the Study:

• To develop a reader-adaptive network (ReaderAdaptNet) that explicitly models inter-reader variability.
• To improve the reliability and personalization of AI-based breast imaging analysis.

Main Methods:

• Proposed a novel two-stage deep learning framework, ReaderAdaptNet, utilizing reader-specific embeddings.
• The first stage learns image features and reader annotation styles; the second stage allows for embedding calibration for rapid adaptation.
• Evaluated on breast density and BPE classification tasks using multi-reader datasets.

Main Results:

• Reader embeddings significantly improved mean classification accuracy: from 76.4% to 84.4% for breast density and 65.1% to 72.1% for BPE.
• Calibrated embeddings enabled flexible, low-cost personalization without full model retraining.
• Demonstrated improved individual and consensus-level performance.

Conclusions:

• ReaderAdaptNet effectively disentangles stable image features from reader-specific decision tendencies.
• Offers a parameter-efficient and interpretable approach for personalization or unification in breast imaging analysis.
• Addresses real-world variability in AI model development.