Customized T-time inner sampling network with uncertainty-aware data augmentation strategy for multi-annotated lesion...

Area of Science:

• Medical Imaging
• Computational Pathology
• Artificial Intelligence in Medicine

Background:

• Medical image segmentation is challenging due to inherent ambiguity.
• Accurately capturing segmentation uncertainty is vital for cancer diagnosis and treatment planning.
• Existing methods for handling segmentation uncertainty are often inefficient and underutilize multi-annotated data.

Purpose of the Study:

• To develop efficient methods for medical image segmentation that leverage multi-annotated datasets.
• To address the limitations of current models in terms of efficiency and utilization of uncertainty information.
• To improve the accuracy and efficiency of lesion segmentation in medical imaging.

Main Methods:

• Customized T-time Inner Sampling Network for flexible sample generation aligned with annotator distributions.
• Definition of Uncertainty Degree for quantitative measurement of sample and dataset uncertainty.
• Uncertainty-aware Data Augmentation Strategy to adapt probabilistic models to varying uncertainty levels.

Main Results:

• Proposed methods demonstrate superior performance in both accuracy and efficiency on lung nodule and liver tumor datasets.
• The Customized T-time Inner Sampling Network efficiently generates samples reflecting ground-truth distribution.
• Uncertainty Degree provides a novel perspective for quantifying segmentation uncertainty and dataset imbalance.

Conclusions:

• The developed methods offer a significant advancement in lesion segmentation by effectively managing uncertainty.
• The approach shows great potential for improving downstream tasks in real-world clinical scenarios.
• This work highlights the importance of leveraging multi-annotated data and uncertainty quantification for robust medical image analysis.