277 Non-invasive imaging biomarkers 3 the risk of malignancy in indeterminate thyroid nodules by combining the ATA classification with cytological subcategorisation (nuclear atypia, architectural atypia, oncocytic atypia) [400]. They found that the risk of malignancy reached almost 80% when both nuclear atypia and ATA-based high-risk US features are present. The presence of these cytological features also increased the risk of malignancy in the ATA-based intermediate-risk category. Architectural atypia and oncocytic patterns were not independently related to higher cancer risk. Moreover, a recent meta-analysis by Staibano et al. including 17 studies investigated sonographic risk criteria (ACR TI-RADS, EU TI-RADS, K-TI-RADS, or ATA) for further prognostication of Bethesda III and IV nodules [401]. In both Bethesda III and Bethesda IV nodules separately, ATA had the highest pooled specificity of 90% and 94% (sensitivity of 52% and 15%), while K-TI-RADS had the highest pooled sensitivity of 78% and 91% (specificity of 53% and 40%), respectively. EU-TI-RADS does not contribute to the clinical management of patients with cytologically indeterminate Hürthle cell nodules, particularly those classified as Bethesda IV [402]. These results underline the combination of cytological subcategorisation and US risk stratification in the management of indeterminate nodules. A conservative approach is proposed in nodules with low-risk US suspicion and Bethesda III, while additional diagnostics and surgery should be considered for nodules with high-risk US suspicion and Bethesda IV or V [276, 403]. AI has been investigated for the optimisation of the ACR TI-RADS risk stratification. Wildman-Tobriner et al. developed AI TI-RADS as a simplification of ACR TI-RADS in unselected nodules, where six features were assigned zero points, using a genetic algorithm inspired by natural selection and its genetic underpinnings [404]. The model was trained using 1,325 nodules and validated using 100 nodules, resulting in similar AUCs for ACR TI-RADS and AI TI-RADS of 91% and 93%, respectively. Specificity of AI TI-RADS (65%) was higher than that of ACR TI-RADS (47%). US radiomics is gaining interest in thyroid nodules. Yoon et al. built a US radiomic score for the differentiation of benign and malignant lesions, retrospectively including 155 nodules with Bethesda III and IV indeterminate cytology [405]. Seven hundred thirty radiomic features were extracted from a square region of interest delineated on a representative 2D image of the initial US. A radiomic score incorporating fifteen radiomic features combined with clinical variables (nodule size, gender, age, Bethesda category) performed significantly better than a model composed of clinical variables only with cross-validated AUCs of 84% and 58%, respectively. Major limitations of this study are the use of clinical US images instead of quantitative images and the choice of a representative image by a human reader. Although inherent to US imaging, bias is introduced by the implicit radiologist input in the selection of the 2D slices as described as the Clever Hans effect by Wallis et al. in a widely used MRI dataset [406]. In addition, in unselected nodules, US radiomic analysis has been extensively studied for the differentiation of benign and malignant nodules. A recent meta-analysis by Cleere et al. including 75 studies found a pooled sensitivity of 87% and a pooled specificity of 84%, which indicates that,
RkJQdWJsaXNoZXIy MTk4NDMw