(Sunday, 7/29/2018) 2:05 PM - 3:00 PM
Room: Room 207
Purpose: Manual detection of bone abnormalities in medical images is time consuming. In this work, we developed a tool to assist in identifying bone abnormalities by providing expected ranges of healthy bone intensity values at a voxel level and applied these ranges to provide the statistical likelihood that each voxel is abnormal.
Methods: Seventeen prostate cancer patients received ¹�F-NaF PET/CT scans. All benign or malignant bone lesions were identified and segmented on NaF PET/CT images by an experienced nuclear medicine physician. A subject was selected from the population to serve as a template, and all healthy portions of each patient’s images were registered to the template using a combination of articulated and deformable registrations. PET images were log-transformed and combined into statistical appearance models, which represented voxel-level averages and standard deviations of healthy bone intensity values. Voxel-wise Z-score images of each patient were created by subtracting the population-averaged image from the patient’s image and then dividing by the population standard deviation image. ROC analysis of Z-scores in the bone was performed to assess the classification performance of Z-scores for detection of abnormal bone.
Results: Of the 447 benign or malignant lesions segmented by the nuclear medicine physician, 312/447 (70%) had Z-score>2 (2 standard deviations above the population mean). 26/32 metastases had Z-score>2. ROC analysis of Z-scores resulted in an area under the curve (AUC) of 0.84. Z-score>3.5 resulted in a sensitivity and specificity of 72% and 78%, respectively, for detecting benign or malignant bone lesions. The Z-score classification performance was primarily degraded by detection of areas of inflammation that had elevated NaF uptake.
Conclusion: We have developed a tool that creates whole-skeleton statistical appearance models of bone. When applied to perform Z-score based detection in NaF PET/CT images, Z-scores accurately detected various abnormalities in the bone.
Funding Support, Disclosures, and Conflict of Interest: Portions of this work are included in a pending patent owned by the Wisconsin Alumni Research Foundation.
Keywords
Image Analysis, Registration, Computer Vision
Taxonomy
IM/TH- Image Analysis (Single modality or Multi-modality): Computer/machine vision
Contact Email
