Program Information
An Optimized Machine Learning Algorithm for the Automated Classification of Benign and Malignant Lesions in [F-18]-NaF PET/CT Images
T Perk1*, T Bradshaw1 , S Chen2 , H Im1 , S Cho1 , S Perlman1 , G Liu1 , R Jeraj1,3 , (1) University of Wisconsin, Madison, WI (2) The 1st hospital of China Medical University, Shenyang, Liaoning (3) University of Ljubljana, Ljubljana, Slovenia
Presentations
SU-K-FS4-1 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: Four Seasons 4
Purpose: 18F-NaF PET/CT imaging of bone metastases is confounded by false positives, due to uptake in benign disease such as osteoarthritis. This work optimizes the performance of an automated bone lesion classification algorithm to accurately classify lesions based on different physician-derived ground truths.
Methods: A nuclear medicine physician analyzed NaF PET/CT scans from thirty-eight metastatic castrate-resistant prostate cancer patients, fourteen of which were analyzed by three additional physicians. All lesions were classified on a five-point scale from definite benign to definite metastatic lesions. Classification agreement between physicians was assessed using Fleiss' κ. ROIs were segmented with three different thresholding methods. For each ROI in the image, 170 different imaging features were extracted, including PET and CT texture and spatial probability features. These imaging features were used as inputs into different machine learning algorithms. Impact of different deterministic factors affecting classification performance, including using different physician-derived ground truths, thresholding methods, algorithm hyperparameters, and imaging resolution used for texture extraction, was assessed.
Results: The factors that most impacted classification performance were the machine learning algorithm and the thresholding method for lesion identification and segmentation. Random forests (RF) had the highest classification performance. Using optimized bone-specific thresholds (AUC=0.95[95%CI=0.94-0.95], sensitivity=85%[84%-86%], and specificity=89%[88%-91%]) resulted in superior performance compared to global thresholds found in literature, SUV>10 g/mL (AUC=0.87) and SUV>15 g/mL (AUC=0.86). There was only moderate agreement between physicians in lesion classification (κ=0.53[95%CI=0.52-0.53]). Classification performance was high using any of the four physicians as ground truth (AUC range:0.91-0.93). Additionally, classification results were superior (AUC=0.95 vs. AUC=0.92) when only using ROIs classified as definitely benign or definitely malignant. Other factors had minimal impact on classification performance.
Conclusion: Our optimized RF algorithm is able to accurately replicate how each physician classifies benign and malignant bone lesions in NaF PET/CT images, allowing for efficient and consistent automated lesion classification.
Funding Support, Disclosures, and Conflict of Interest: This project was funded by the PCF challenge.
Contact Email: