359 Quantitative classification and radiomics of [18F]FDG-PET/CT 5 unsupervised feature selection or dimensionality reduction over supervised feature selection, which uses discriminative values for the outcome. Unsupervised methods take into account the interaction of features and multicollinearity, thereby preventing overfitting of the model [521]. We selected nonredundant features with low multicollinearity, which were not necessarily the features with the highest predictive performance. Second, dimensionality reduction was performed on the training sets in the folds instead of on the dataset as a whole, strictly distinguishing the independent test sets. Third, factor-based dimensionality reduction was chosen over a feature-based approach for generalizability purposes. Instead of selecting features corresponding to the retained factors, the factors were used as input for the model and patterns in corresponding features were compared between folds. In a featurebased approach, different features might have been selected in different folds, resulting in limited insight in these patterns. Along these lines, a factor-based approach improves the generalizability and interpretability of the model and might provide insight in the semantics or underlying tumour biology of the factors [393]. Contrarily, it reduces the (mathematical) explainability and reproducibility of the radiomic model during external validation, as it uses derivatives of features. Adherence to the IBSI reporting guidelines and TRIPOD statement may prevent reproducibility issues [514, 518]. Another limitation is that eighteen nodules did not meet the minimal size recommendation for radiomic analysis of 64 voxels per VOI [517]. Subgroup analysis of the nodules meeting this requirement showed similar results. It is unlikely that the nodule size had a large impact on the radiomic analysis. The multicentre design of the study was both a strength and limitation. While the population of our nationwide trial is unique and an adequate reflection of the diverse presentation of thyroid nodules, the different scanners and slight variations in imaging protocols among the 12 hospitals introduced heterogeneity and may have limited the radiomic analysis. Therefore, only scans with strict adherence to the EANM guidelines were assessed, as these reconstructions leads to a larger number of reliable, repeatable, and reproducible radiomic features in a multicentre and multivendor setting [522]. In addition, nodules were delineated using a threshold of 50% of the SUVpeak, corrected for local background, which is recommended in multicentre [18F]FDG PET/CT studies because of its high feasibility and repeatability [511]. Moreover, all images were interpolated to isotropic voxels in order to allow comparison between image data from different samples and centres [523]. The number of included patients per centre or PET/CT-scanner was not sufficiently large to incorporate post-reconstruction harmonization strategies such as ComBat [524]. In conclusion, the current study showed that quantitative [18F]FDG-PET/CT assessment accurately ruled out malignancy in both Hürthle cell and non-Hürthle cell indeterminate thyroid nodules. Distinctive SUV cut-offs may avoid up to one in three futile diagnostic surgeries for benign Hürthle cell nodules. In non-Hürthle cell nodules, quantitative assessment had no added diagnostic value over visual [18F]FDG-PET/CT assessment. Radiomic analysis did not contribute to the additional differentiation of [18F]FDG-positive thyroid nodules in this dataset.
RkJQdWJsaXNoZXIy MTk4NDMw