Quantifying central canal stenosis prediction uncertainty in SpineNet with conformal prediction

Area of Science:

• Medical Imaging
• Machine Learning
• Radiology

Background:

• Central canal stenosis (CCS) grading from MRI is crucial for patient management.
• Accurate prediction of CCS severity is essential for clinical decision-making.
• Quantifying uncertainty in AI model predictions is vital for clinical adoption.

Purpose of the Study:

• To apply conformal prediction (CP) methods to SpineNet for quantifying uncertainty in CCS classification.
• To evaluate the performance of different CP techniques in grading CCS severity.
• To identify the most reliable CP method for clinical application in CCS assessment.

Main Methods:

• Analysis of 1689 vertebral levels from 340 patients undergoing T2-weighted MRI.
• Application and evaluation of four conformal prediction methods: class-conditional CP, Top-k, Least Ambiguous Set-Valued Classifiers (LAC), and Adaptive Prediction Sets (APS).
• Bootstrap resampling used for robustness assessment across calibration/test splits; multiple significance levels (α) evaluated.

Main Results:

• Class-conditional CP consistently achieved desired coverage with the smallest prediction sets.
• Top-k, LAC, and APS methods showed limitations, producing larger or less informative prediction sets, especially for moderate/severe CCS.
• Class-conditional CP demonstrated class- and level-specific uncertainty, with larger sets for moderate/severe grades, indicating higher uncertainty.

Conclusions:

• Class-conditional CP is the most reliable and clinically informative method for estimating uncertainty in CCS grading.
• CP provides a transparent approach to assess SpineNet's reliability in medical predictions.
• This approach enhances decision-making by providing quantified uncertainty alongside CCS severity classification.