Quality Assurance using Outlier Detection on an Automatic Segmentation Method for the Cerebellar Peduncles

Area of Science:

• Neuroimaging
• Medical Image Analysis
• Computational Neuroscience

Background:

• Cerebellar peduncles (CPs) are vital white matter tracts connecting the cerebellum to other brain regions.
• Automatic segmentation of CPs aids in studying their structure and function.
• Evaluating automatic segmentation performance typically relies on manual delineations (ground truth), which are unavailable for new datasets.

Purpose of the Study:

• To develop and evaluate methods for automatically detecting failures in cerebellar peduncle segmentation algorithms.
• To enable efficient exclusion of erroneous segmentation results from scientific analysis when ground truth is absent.

Main Methods:

• Two outlier detection methods were investigated: a univariate non-parametric approach using box-whisker plots and supervised classification.
• Features were designed from diffusion tensor imaging (DTI) data to categorize segmentation results as success or failure.
• Four classification algorithms (LDA, LR, SVM, RFC) were trained and evaluated using leave-one-out cross-validation on a dataset of 249 subjects.

Main Results:

• The univariate method demonstrated that designed features could efficiently detect true segmentation failures.
• Supervised classification methods, particularly LDA, SVM, and RFC, showed comparable performance in automatically identifying segmentation failures.
• Logistic regression (LR) performed the worst among the evaluated classifiers.

Conclusions:

• Automatic detection of cerebellar peduncle segmentation failures is feasible using feature-based outlier detection and classification techniques.
• These methods can reliably assess the performance of automatic segmentation algorithms on new DTI data without requiring manual ground truth.
• The developed techniques facilitate more robust and accurate analysis of cerebellar peduncle structure and function in neuroimaging studies.