Impact of deep learning based reconstruction algorithms on CT radiomic features of carotid plaques

TL;DR

This study examines how different CT image reconstruction methods affect the reliability of radiomic features in carotid plaques, finding that texture features are more stable than others.

Contribution

The study quantifies the impact of deep learning and iterative reconstruction algorithms on radiomic feature reproducibility in carotid plaques.

Findings

Texture features are more stable across reconstruction methods compared to first-order features.

Higher-strength DLIR and ASIR-V settings reduce the consistency of radiomic features.

First-order features show excellent inter-observer agreement in 3D plaque analysis.

Abstract

Radiomics is increasingly applied in carotid plaques analysis to evaluate plaque characteristics and predict cardiovascular risk. However, the influence of different image reconstruction algorithms, particularly deep learning reconstruction (DLIR) and adaptive statistical iterative reconstruction‐Veo (ASIR‐V), on the reproducibility of radiomic features remains poorly understood.

To evaluate the impact of DLIR and ASIR‐V on CT radiomic features of carotid plaques.

76 patients with 104 carotid plaques who underwent head & neck CT angiography were retrospectively enrolled. Images were reconstructed by filtered back projection (FBP), ASIR‐V (30%, 50%, and 80%) and DLIR (DL, DM, and DH). A total of 214 CT‐based radiomic features were organized by statistic family (18 first‐order; 75 texture: 24 GLCM, 14 GLDM, 16 GLRLM, 16 GLSZM, and 5 NGTDM) and transform domain (original and wavelet sub‐bands); 121 features were extracted from wavelet sub‐bands. Features were extracted from both 2D and 3D plaque images. The reliability of feature extraction was evaluated by the intraclass correlation coefficient (ICC).

Different reconstruction algorithms influenced the most radiomic features. The percentages of first‐order, texture, and features in the wavelet domain without statistical difference among 2D and 3D lesions for all seven groups were 0% (0/18), 12.0% (9/75), and 14.9% (18/121), respectively. Compared with FBP, the unaffected features for AV30%, 50%, and 80% decreased from 99.8% and 95.1% to 81.3%, and for DL, DM, and DH from 75.5% and 52.3% to 40.7%. Across statistic families, texture features were the most stable in pairwise comparisons in both the original and wavelet domains. Unaffected features in 2D lesion were larger than 3D lesion. The consistency of first‐order feature in 3D lesion was excellent in both intra‐ and inter‐observer, with ICC values ranging from 0.865 to 1 and 0.790 to 0.999, respectively.

Both ASIR‐V and DLIR algorithms profoundly impact carotid plaque radiomics, with higher strengths exacerbating feature instability. Texture features exhibited superior robustness across all reconstruction protocols. Our findings advocate for a stability‐driven approach to model development: prioritizing robust texture features and employing lower‐strength DLIR are crucial steps to ensure the generalizability of radiomic biomarkers.

Figures