Propagating Uncertainty Across Cascaded Medical Imaging Tasks for Improved Deep Learning Inference

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Pathology

Background:

• Deep networks excel in medical imaging but struggle with pathology-induced challenges.
• Cascading deterministic outputs in clinical workflows leads to cumulative errors, hindering downstream task accuracy.

Purpose of the Study:

• To improve deep learning model performance in clinical workflows by embedding uncertainty estimates across cascaded inference tasks.
• To demonstrate the effectiveness of uncertainty propagation in enhancing downstream medical imaging analysis.

Main Methods:

• Propagating uncertainty estimates from T2 weighted lesion segmentation to improve T2 lesion detection.
• Incorporating uncertainty maps from synthesized missing MR volumes to enhance brain tumor segmentation.
• Propagating uncertainties from hippocampus segmentation for improved Alzheimer's disease clinical score regression.

Main Results:

• Improved T2 lesion detection performance on a large-scale, multi-site Multiple Sclerosis dataset.
• Enhanced brain tumor segmentation accuracy using synthesized missing MR volume uncertainty maps on the BraTS-2018 dataset.
• Improved regression of Alzheimer's disease clinical scores through propagated voxel-level hippocampus segmentation uncertainties.

Conclusions:

• Embedding uncertainty estimates across cascaded inference tasks significantly improves performance in downstream medical imaging applications.
• The proposed uncertainty propagation method offers a robust solution to enhance the reliability and accuracy of deep learning models in clinical settings.
• This approach addresses key challenges in integrating deep learning into clinical workflows, paving the way for more reliable AI-driven diagnostics.