Evaluating the robustness of features generated by a foundation model from CT with different reconstruction parameters

评估基于不同重建参数的CT基础模型所生成特征的鲁棒性

阅读:4

Abstract

Foundation models are increasingly recognized for their ability to extract robust imaging features by learning generalized representations from large datasets of unannotated images. Foundation models show promise in generating task-agnostic features that outperform conventional methods in diagnostic and prognostic tasks. However, the clinical application of imaging biomarkers is often hampered by their sensitivity to variations in image acquisition and reconstruction parameters. Hence, we evaluated the robustness of features extracted using the pretrained foundation model from low-dose computed tomography (LDCT) scans under varying image conditions. We analyzed 59 LDCT scans containing at least one nodule from individuals who underwent lung cancer screening. Each scan consisted of 6 variations of image conditions, including combinations of three reconstruction kernels (smooth, medium, sharp) and two slice thicknesses (1.0 mm, 2.0 mm). 56 nodules had nodule outcome labels (cancer vs. non-cancer) available, of which 7 had cancer outcomes. Features were extracted using the foundation model across all image conditions. The feature-level robustness was evaluated by computing the concordance correlation coefficient (CCC) between each feature value extracted from the reference condition (Medium/1.0) and the other 5 conditions. We also examined the impact of varying image conditions on the features' predictive ability in nodule outcome classification. Pyradiomics, a well-established radiomic feature extractor, served as a baseline for comparison. The foundation model consistently yielded high mean CCC values ranging from 0.937 (Smooth/2.0) to 0.984 (Sharp/1.0), indicating robustness against variations in image conditions. Moreover, the foundation model's features demonstrated high performance in nodule outcome classification compared to Pyradiomics features. However, inconsistencies in performance across different image conditions underscore the need for further exploration of harmonization techniques and modeling strategies to enhance the generalizability and clinical utility of imaging biomarkers.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。