人工智能驱动的全身[18F]FDG PET/CT图像质量评估系统
本研究纳入173例全身[18F]FDG PET/CT患者图像,将163例重建为952幅图像并随机分为训练集(98例578幅)、验证集(33例192幅)和测试集(32例182幅),开发卷积神经网络图像质量评估系统(IQA-CNNs)。IQA-CNNs与医师主观评估的kappa系数为0.79和0.78,客观评估ICC>0.60,评估速度比人工快200倍,性能与高年资医师相当。
要点
- IQA-CNNs与R1、R2主观评估的kappa系数分别为0.79和0.78
- IQA-CNNs与R1或R2客观图像质量评估ICC>0.60
- IQA-CNNs评估速度比人工评估快200倍
核心结论
- 开发了快速评估[18F]FDG PET/CT图像质量的自动化系统,性能与高年资医师相当。
- 该系统生成包含各解剖区域主观视觉评分和客观图像指标的全面详细图像质量评估报告。
研究设计与样本
- 本研究纳入173例(年龄59±12岁,66.3%为男性)行全身[18F]FDG PET/CT显像的患者。
- 10例患者图像作为教学集,其余163例患者图像通过模拟不同扫描时间重建为952幅图像。
- 952幅图像随机分为训练集(60%,98例,578幅)、验证集(20%,33例,192幅)和测试集(20%,32例,182幅)。
- 两名经验丰富的医师(R1和R2)使用5点Likert量表独立评估胸部、腹部和盆腔区域的图像质量两次(R1a和b;R2a和b),间隔1个月。
- 从纵隔血池、三个肝脏层面和双侧臀大肌提取客观图像质量指标,开发卷积神经网络图像质量评估系统(IQA-CNNs)。
主要发现
- 医师间和医师内主观评估一致性良好,kappa系数(R1a vs. R2a、R1a vs. R1b、R2a vs. R2b、R1a vs. R2b)分别为0.78、0.77、0.76和0.80。
- IQA-CNNs与R1或R2主观评估一致,kappa系数分别为0.79和0.78。
- IQA-CNNs与R1或R2客观图像质量评估一致(ICC>0.60)。
- IQA-CNNs评估速度比人工评估快200倍。
关键数值与推荐
- 医师间主观评估kappa系数:R1a vs. R2a为0.78,R2a vs. R2b为0.76。
- 医师内主观评估kappa系数:R1a vs. R1b为0.77,R1a vs. R2b为0.80。
- IQA-CNNs与R1主观评估kappa系数为0.79,与R2为0.78。
- IQA-CNNs与R1或R2客观图像质量评估ICC>0.60。
局限性
- 原文未报告局限性
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
PURPOSE: Image quality control is a prerequisite for applying PET/CT. This study aimed to develop an artificial intelligence-driven real-time and accurate whole-body [18F]FDG PET/CT image quality assessment system. METHODS: This study included 173 patients (age, 59 ± 12 years; 66.3% males) with whole-body [18F]FDG PET/CT imaging. Images of ten patients were used as an educational set. Images of the rest 163 patients were reconstructed to 952 images by simulating several scanning times and randomly split into training (60%, 98 patients, 578 images), validation (20%, 33 patients, 192 images), and test (20%, 32 patients,182 images) sets. Two experienced physicians (R1 and R2) independently assessed the image quality of thorax, abdomen, and pelvis region twice (R1a and b; R2a and b), 1 month apart, using a 5-point Likert scale. Objective image quality metrics were extracted from the mediastinal blood pool, three liver levels, and the bilateral gluteus maximus. The developed convolutional neural networks for image quality assessment (IQA-CNNs) generated the subjective quality scores and objective image metrics. The IQA-CNNs and physicians' performances were compared for localization accuracy, score agreement, and process time. RESULTS: The physicians demonstrated good inter- and intra-rater subjective assessment agreement, with kappa coefficients (R1a vs. R2a, R1a vs. R1b, R2a vs. R2b, and R1a vs. R2b) of 0.78, 0.77, 0.76, and 0.80. The IQA-CNNs and R1 or R2 agreed in the subjective assessments, with kappa coefficients of 0.79 and 0.78. IQA-CNNs and R1 or R2 also agreed in their objective image quality assessment (ICC > 0.60). The IQA-CNNs evaluation speed was 200 times faster than the manual assessment. CONCLUSION: An automated system for rapid assessment of [18F]FDG PET/CT image quality was developed, showing comparable performance to senior physicians. The system generates a comprehensive and detailed image quality assessment report, including subjective visual scores and objective image metrics for various anatomical regions.
引用
- Qi C, Wang S, Yu H, et al. An artificial intelligence-driven image quality assessment system for whole-body [18F]FDG PET/CT. Eur J Nucl Med Mol Imaging. 2022. PMID: 36529840. doi:10.1007/s00259-022-06078-z
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。