The effect of variable labels on deep learning models trained to predict breast density.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Biostatistics

Background:

• High breast density is linked to reduced mammographic screening efficacy and increased breast cancer risk.
• Automated density estimates offer potential for direct risk prediction and integration into predictive models.
• Expert reader assessments of breast density correlate with cancer risk but exhibit inter-reader variability, impacting model development.

Purpose of the Study:

• To investigate the effect of expert reader label variability on automated mammographic density estimation models.
• To assess how label variability influences the mapping from model representation to prediction.
• To evaluate the impact of label variability on the learned model representation for density prediction.

Main Methods:

• Utilized subsets of mammographic images with density labels from 13 readers and 12 reader pairs.
• Trained a deep transfer learning model to assess label variability's effect on representation-to-prediction mapping.
• Developed two end-to-end models: one trained on averaged labels, another on individual reader scores with modified objective function.

Main Results:

• Trained mappings from representations to labels were considerably altered by reader score variability.
• Removing distribution variation in labels increased Spearman rank correlation coefficients from 0.751 to 0.815 (averaged readers) or 0.844 (averaged images).
• Investigating representation effects showed little difference, with statistically similar Spearman rank correlation coefficients (0.846-0.850) regardless of label variability.

Conclusions:

• Mammographic density prediction models are significantly affected by label variability in the mapping between representation and prediction.
• The underlying model representation's quality for density prediction is minimally affected by label variability.
• Findings highlight the importance of addressing label variability for accurate automated breast density assessment and risk prediction.