AI去噪在儿童胸部X线模拟减剂量中的前瞻性多中心评估
这项前瞻性多中心研究纳入420例儿童胸部X线，用物理模型模拟约50%减剂量并施加深度学习去噪。AI去噪使主观图像质量接近全剂量（4.4±0.5 vs 4.5±0.4），诊断性能绝对差异小（ΔAUC=-0.03），但置信区间跨越预设非劣效界值，未正式确立非劣效。
要点
- AI去噪可改善模拟50%减剂量儿童胸片的主观图像质量。
- 诊断性能ΔAUC=-0.03，但CI跨越-0.05非劣效界值，非劣效未成立。
- 临床等效性尚不能假定，需真实低剂量数据进一步验证。
核心结论
- AI去噪改善模拟低剂量儿童胸片图像质量，但诊断性能未达正式非劣效。
研究设计与样本
- 前瞻性多中心研究，纳入420例临床需行胸部X线的儿童患者，来自四家三级医院。
- 使用经验证的物理信息泊松-高斯噪声插入框架生成模拟低剂量图像（全部患者50%，探索性70%亚组）。
- 图像经深度残差卷积神经网络处理，由三名盲法儿童放射科医师独立评估图像质量及对肺炎、肺不张、心脏增大的诊断性能。
- 采用定量指标和ROC分析（DeLong法），预设ΔAUC非劣效界值为-0.05。
诊断标准与关键数值
- 预设非劣效界值为ΔAUC -0.05，实际ΔAUC=-0.03，但置信区间跨越该界值，非劣效未成立。
- AI去噪后主观图像质量平均评分4.4±0.5，接近全剂量4.5±0.4。
临床应用与阅片要点
- AI去噪有望用于儿童胸片减剂量，但当前证据尚不足以认定诊断等效。
- 报告时应关注模拟低剂量与真实低剂量的差异，避免直接外推至临床。
局限性
- 采用模拟低剂量而非临床实际采集的低剂量数据。
- 诊断性能未达预设非劣效，临床等效性不能假定。
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）。
英文原文（摘要）· 点击展开
RATIONALE AND OBJECTIVES: Radiation dose reduction in pediatric chest radiography is a clinical priority due to increased radiosensitivity and cumulative lifetime risk. However, substantial dose reduction may degrade image quality and compromise diagnostic performance. Artificial intelligence (AI)-based denoising has emerged as a potential solution, although robust prospective multicenter validation remains limited. To prospectively evaluate whether a vendor-neutral deep learning-based denoising system improves image quality while assessing the magnitude of change in diagnostic performance in simulated low-dose pediatric chest radiographs at approximately 50% dose reduction.
MATERIALS AND METHODS: In this prospective multicenter study, 420 pediatric patients undergoing clinically indicated chest radiography were enrolled across four tertiary hospitals. Simulated low-dose images (50% for all patients; exploratory 70% subset) were generated using a validated physics-informed Poisson-Gaussian noise-insertion framework. Images were processed using a deep residual convolutional neural network (CNN). Three blinded pediatric radiologists independently assessed image quality and diagnostic performance for pneumonia, atelectasis, and cardiomegaly. Quantitative metrics and receiver operating characteristic (ROC) analysis with DeLong's method were performed. Non-inferiority was tested using a prespecified ΔAUC margin of -0.05.
RESULTS: AI denoising improved perceived image quality, with mean scores approaching those of full-dose imaging (4.4 ± 0.5 vs 4.5 ± 0.4). Quantitative metrics demonstrated preservation of structural similarity and noise characteristics. Although the absolute difference in diagnostic performance was small (ΔAUC = -0.03), the confidence interval (CI) crossed the prespecified non-inferiority margin; therefore, formal non-inferiority was not established.
CONCLUSION: AI-based denoising improves image quality in simulated low-dose pediatric chest radiography. Although the absolute difference in diagnostic performance was small (ΔAUC = -0.03), the CI crossed the prespecified non-inferiority margin; therefore, formal non-inferiority was not established. These findings support further prospective validation using clinically acquired low-dose data before clinical equivalence can be assumed.
引用
- Shakhreet BZ. AI-Based Denoising for Simulated Dose Reduction in Pediatric Chest Radiography: A Prospective Multicenter Evaluation of Image Quality and Diagnostic Performance. Acad Radiol. 2026. PMID: 42034503. doi:10.1016/j.acra.2026.04.003
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。