Deep Quantile Regression for Uncertainty Estimation in Unsupervised and Supervised Lesion Detection.

Area of Science:

• Machine Learning
• Medical Image Analysis
• Deep Learning

Background:

• Deep learning models can be over-confident, especially with limited data.
• Quantifying uncertainty is critical for medical applications like lesion detection and diagnosis.
• Existing Variational AutoEncoder (VAE) methods suffer from variance shrinkage.

Purpose of the Study:

• To propose quantile regression for quantifying aleatoric uncertainty in lesion detection.
• To develop unsupervised (QR-VAE) and supervised (BQR) methods for uncertainty quantification.
• To improve the reliability of deep learning in critical medical applications.

Main Methods:

• Quantile regression applied to unsupervised lesion detection using a Variational AutoEncoder (VAE).
• Developed Quantile-Regression VAE (QR-VAE) to avoid variance shrinkage by directly estimating conditional quantiles.
• Developed Binary Quantile Regression (BQR) for supervised lesion segmentation.

Main Results:

• QR-VAE effectively quantifies uncertainty in VAE reconstructions for outlier detection.
• BQR captures uncertainty in lesion boundaries, reflecting expert disagreement.
• Both methods enhance lesion detection and segmentation accuracy.

Conclusions:

• Quantile regression offers a robust approach to aleatoric uncertainty quantification in deep learning for medical imaging.
• The proposed QR-VAE and BQR methods provide more reliable confidence estimates than standard VAEs.
• These advancements are vital for trustworthy clinical diagnosis and treatment planning.