Related Experiment Video
Updated: Jul 7, 2026
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
5.8K
Uncertainty quantification for deep learning-based metastatic lesion segmentation on whole body PET/CT
Brayden Schott1, Victor Santoro-Fernandes2, Zan Klanecek3
1Department of Medical Physics, University of Wisconsin, 1111 Highland Ave #1005, Madison, Wisconsin, 53705, UNITED STATES.
Physics in Medicine and Biology
|May 16, 2025
Summary
Test time augmentation (TTA) is the best uncertainty quantification method for segmenting metastatic lesions on whole body PET/CT scans. Probability entropy performed poorly, indicating a need for advanced uncertainty quantification approaches.
Area of Science:
- Medical Imaging Analysis
- Artificial Intelligence in Healthcare
- Radiopharmaceutical Theranostics
Background:
- Deep learning models are crucial for automated medical image analysis but often lack uncertainty quantification (UQ).
- Ensuring the reliability of model outputs is essential for patient care, especially in complex tasks like metastatic lesion segmentation.
- Existing UQ methods vary in effectiveness, and optimal choices for specific tasks remain unclear.
Purpose of the Study:
- To investigate and compare commonly used UQ methods for metastatic lesion segmentation on whole body PET/CT.
- To evaluate the performance of probability entropy, Monte Carlo dropout, deep ensembles, and test time augmentation (TTA).
- To determine the most suitable UQ method for enhancing the reliability of AI-driven medical image analysis in oncology.
Main Methods:
- Utilized 59 whole body 68Ga-DOTATATE PET/CT images from patients with metastatic neuroendocrine tumors.
- Trained a 3D U-Net model for lesion segmentation using five-fold cross-validation.
- Assessed four UQ methods (probability entropy, Monte Carlo dropout, deep ensembles, TTA) based on image degradation detection, false-positive/negative identification, and correlation with performance metrics.
Main Results:
- Test time augmentation (TTA) demonstrated superior performance in detecting degraded image data across all magnitudes.
- All UQ methods showed strong performance in detecting false-positive regions (AUC 0.77-0.81).
- TTA excelled in false-negative region recovery and showed strong correlations with SUVtotal and Dice coefficient, while probability entropy performed worst.
Conclusions:
- Test time augmentation (TTA) is recommended as a superior UQ method for metastatic lesion segmentation due to its effectiveness and efficiency.
- Probability entropy was found to be the least effective UQ method for this task.
- The study highlights the critical need for advanced UQ approaches to ensure the reliability of deep learning models in medical imaging.
