Task-based evaluation of segmentation algorithms for diffusion-weighted MRI without using a gold standard.

Area of Science:

• Medical Imaging
• Radiology
• Biomedical Engineering

Background:

• Accurate apparent diffusion coefficient (ADC) estimation from diffusion-weighted (DW) MRI is crucial for visceral organ lesion characterization.
• Current methods for evaluating lesion segmentation algorithms, like region-overlap measures, do not directly assess the primary task of ADC estimation and often require unavailable gold-standard segmentations.

Purpose of the Study:

• To address the limitations of current evaluation methods for lesion segmentation algorithms in DW imaging.
• To develop and validate a task-based evaluation framework that does not rely on gold-standard or manual segmentations.
• To enable reliable comparison and ranking of segmentation algorithms based on their impact on ADC estimation accuracy.

Main Methods:

• Investigated the unreliability of using manual segmentations for task-based evaluation in the absence of a gold standard.
• Proposed a novel no-gold-standard method to estimate the bias and variance of ADC estimation errors.
• Developed consistency checks to validate the proposed evaluation technique.

Main Results:

• Demonstrated that manual segmentations are unreliable for task-based evaluation when a gold standard is absent.
• The proposed no-gold-standard method effectively estimates bias and variance of ADC estimation errors.
• The method allows for ranking segmentation algorithms based on ensemble mean square error and precision.

Conclusions:

• A reliable, gold-standard-free method for task-based evaluation of lesion segmentation algorithms in DW imaging has been developed.
• This approach directly assesses the impact of segmentation on ADC estimation, offering a more relevant evaluation than region-overlap measures.
• The proposed method and consistency checks provide a robust framework for advancing automated segmentation in medical imaging.