Extract
We thank B. Zhang and S. Zhang for their interest in our recently published article [1]. Radiomics is a quite active field of research in computational medicine with multiple studies discussing their limitations and perspectives. As B. Zhang and S. Zhang suggested in their letter, generalisation of machine-learning based models is an open issue, which has been investigated for radiomics models but also for deep-learning models [2]. In contrast, we believe that multicentre data acquired on 14 different computed tomography (CT) scanners with varied acquisition protocols could be considered as representing real-world practice, which is a key element for generalisability. The best way to assess whether radiomic features are reproducible is to repeat the entire feature extraction process, including the CT acquisition. However, repeating CT acquisition is not feasible in a retrospective study and raises ethical concerns in a population frequently exposed from childhood to ionising radiation from medical procedures [3, 4]. In our study, all CT images were normalised before the radiomic analysis in order to avoid bias. This process included preprocessing in 1 mm spatial resolution and the activation of the normalisation function of pyradiomics, which normalised the image densities by centring it at the mean with standard deviation.
抽象的
The diversity of possible approaches in artificial intelligence is the seedbed for new discoveries in thoracic imaginghttps://bit.ly/3A4yyYu
脚注
利益冲突:P-R。Burgel报告支持Vaincre la Mucoviscidose目前的手稿;来自顶点和史克的赠款;来自阿斯利康,基斯,gsk,insmed,insmed,vertex和zambon的咨询费;辉瑞和诺华的演讲酬金;在提交的工作之外。G. Chassagnon在提交的作品之外报告了Chiesi的演讲Honararia。所有其他作者都无话可说。
- 已收到January 5, 2022.
- Accepted2022年1月7日。
- 版权所有©作者2022。用于复制权和权限联系权限{at} ersnet.org