Practical no-gold-standard evaluation framework for quantitative imaging methods: application to lesion segmentation...

Area of Science:

• Medical Imaging
• Quantitative Imaging Analysis
• Positron Emission Tomography (PET)

Background:

• No-gold-standard (NGS) techniques offer a way to assess quantitative imaging precision without gold standards or repeated measurements.
• Practical application of NGS methods to patient data faces challenges in assumption assessment, sampling uncertainty, and reliability evaluation.
• Evaluating lesion segmentation in F-Fluoro-2-deoxyglucose (FDG) PET for head-and-neck cancer requires precise metabolic tumor volume measurement.

Purpose of the Study:

• To develop statistical tests and a bootstrap-based methodology to enhance the reliability and address practical difficulties of NGS techniques.
• To provide confidence in the underlying assumptions and the reliability of estimated figures of merit (FoMs) from NGS.
• To evaluate four lesion segmentation methods for metabolic tumor volume measurement in head-and-neck cancer using FDG-PET data.

Main Methods:

• Proposed statistical tests to validate NGS assumptions and the reliability of estimated FoMs.
• Integrated NGS with a bootstrap-based methodology to account for patient-sampling uncertainty.
• Applied the developed NGS framework to segment lesions from FDG-PET images of head-and-neck cancer patients.

Main Results:

• The NGS technique consistently identified the same segmentation method as the most precise.
• The proposed framework provided reliable results even without gold-standard data.
• Bootstrap analysis showed improved NGS performance with more patient studies and consistent results with over 80 lesions.

Conclusions:

• The developed NGS framework enhances the reliability of quantitative imaging method evaluation using patient data.
• The methodology provides confidence in assessing imaging precision, particularly for metabolic tumor volume in head-and-neck cancer.
• The approach is robust and improves with larger datasets, offering a valuable tool for quantitative imaging research.