Explainable AI for automated respiratory misalignment detection in PET/CT imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Positron emission tomography (PET) and computed tomography (CT) image quality is crucial for diagnosis.
• Misalignment artifacts between PET and CT images can negatively impact diagnostic accuracy.
• Automated detection of these artifacts is needed for efficient data curation and clinical workflow.

Purpose of the Study:

• To develop an explainable machine learning (ML) approach for detecting respiratory misalignment artifacts (RMA) in PET/CT imaging.
• To create an automated pipeline for segmenting organs in PET and CT images and comparing them to identify misalignment.
• To enhance the reliability and efficiency of PET/CT data analysis.

Main Methods:

• Utilized a dataset of 1216 PET/CT images.
• Developed an ML model using a random forest framework with 10-fold cross-validation.
• Compared segmentations of four organs (lungs, liver, spleen, heart) from PET and CT images to detect misalignment.
• Evaluated model performance using sensitivity, specificity, F1-Score, and Area Under the Curve (AUC).

Main Results:

• Achieved high performance metrics: AUC of 0.91 on cross-validation and 0.90 on the test set.
• Demonstrated sensitivity and specificity of 0.82 and 0.85 respectively in cross-validation.
• Identified the liver and lungs as the most significant organs for detecting misalignment.
• Reported F1-Score, sensitivity, and specificity values exceeding 80% for both cross-validation and test sets.

Conclusions:

• An automated, explainable ML pipeline effectively detects misalignment artifacts in PET/CT scans.
• The method mimics human readers by comparing organ segmentations across modalities.
• This approach can significantly improve large dataset curation and can be integrated into clinical scanners for real-time artifact detection.