Fully Automated, Semantic Segmentation of Whole-Body 18 F-FDG PET/CT Images Based on Data-Centric Artificial Intelligence

The Journal of Nuclear Medicine

article

Lalith Kumar Shiyam Sundar¹  Josef Yu¹  Otto Muzik²  Oana C. Kulterer³  Barbara Fueger³  Daria Kifjak³  Thomas Nakuz³  Hyung Min Shin⁴  Annika Katharina Sima³  Daniela Kitzmantl³  Ramsey D. Badawi⁵  Lorenzo Nardo⁵  Simon R. Cherry⁵  Benjamin A. Spencer⁵  Marcus Hacker³  Thomas Beyer¹ 

[1] Quantitative Imaging and Medical Physics Team, Center for Medical Physics and Biomedical Engineering, Medical University of Vienna;Department of Pediatrics, Wayne State University School of Medicine, Children’s Hospital of Michigan;Department of Biomedical Imaging and Image-Guided Therapy, Medical University of Vienna;Division of General Surgery, Department of Surgery, Medical University of Vienna;Department of Biomedical Engineering and Radiology, University of California–Davis

关键词: multiorgan segmentation;    total-body PET;    systems medicine;    artificial neural networks;    automated segmentation;   

DOI  :  10.2967/jnumed.122.264063

学科分类：医学（综合）

来源: Society of Nuclear Medicine

【 摘 要 】

We introduce multiple-organ objective segmentation (MOOSE) software that generates subject-specific, multiorgan segmentation using data-centric artificial intelligence principles to facilitate high-throughput systemic investigations of the human body via whole-body PET imaging. Methods: Image data from 2 PET/CT systems were used in training MOOSE. For noncerebral structures, 50 whole-body CT images were used, 30 of which were acquired from healthy controls (14 men and 16 women), and 20 datasets were acquired from oncology patients (14 men and 6 women). Noncerebral tissues consisted of 13 abdominal organs, 20 bone segments, subcutaneous fat, visceral fat, psoas muscle, and skeletal muscle. An expert panel manually segmented all noncerebral structures except for subcutaneous fat, visceral fat, and skeletal muscle, which were semiautomatically segmented using thresholding. A majority-voting algorithm was used to generate a reference-standard segmentation. From the 50 CT datasets, 40 were used for training and 10 for testing. For cerebral structures, 34 18F-FDG PET/MRI brain image volumes were used from 10 healthy controls (5 men and 5 women imaged twice) and 14 nonlesional epilepsy patients (7 men and 7 women). Only 18F-FDG PET images were considered for training: 24 and 10 of 34 volumes were used for training and testing, respectively. The Dice score coefficient (DSC) was used as the primary metric, and the average symmetric surface distance as a secondary metric, to evaluate the automated segmentation performance. Results: An excellent overlap between the reference labels and MOOSE-derived organ segmentations was observed: 92% of noncerebral tissues showed DSCs of more than 0.90, whereas a few organs exhibited lower DSCs (e.g., adrenal glands [0.72], pancreas [0.85], and bladder [0.86]). The median DSCs of brain subregions derived from PET images were lower. Only 29% of the brain segments had a median DSC of more than 0.90, whereas segmentation of 60% of regions yielded a median DSC of 0.80–0.89. The results of the average symmetric surface distance analysis demonstrated that the average distance between the reference standard and the automatically segmented tissue surfaces (organs, bones, and brain regions) lies within the size of image voxels (2 mm). Conclusion: The proposed segmentation pipeline allows automatic segmentation of 120 unique tissues from whole-body 18F-FDG PET/CT images with high accuracy.

【 授权许可】

CC BY