橋本病における甲状腺結節診断の革新：AI駆動型ラジオミクスおよび深層学習モデル

Accurately distinguishing between benign and malignant thyroid nodules(TNs) in the context of Hashimoto's thyroiditis (HT) is challenging. This study aimed to explore the diagnostic efficacy of artificial intelligence (AI) models constructed using radiomics and deep learning (DL) features extracted from the ultrasound images of TNs in the setting of HT. The study also aimed to quantitatively compare the diagnostic performance of these models against that of fine-needle aspiration (FNA) cytology combined with BRAFV600E gene mutation testing so as to establish a superior diagnostic paradigm for TNs in HT.

We analyzed the clinical data and preoperative ultrasound images of 1585 patients with HT admitted to 8 hospitals in China between 1 January 2018, and 30 December 2023. Radiomics features were extracted from each manually annotated and standardized region of interest in the images. The DL features based on various convolutional neural network models were also extracted, including 11 pre-trained DL models based on real images and the image features extracted using a combined DL-radiomics (DLR) approach. Least absolute shrinkage and selection operator regression was used to select features with nonzero coefficients. Further, eight machine learning methods were employed to construct prediction models, namely the DLR models, incorporating both DL and radiomics features. The importance of features in contributing to the model was prioritized using the SHapley Additive exPlanations (SHAP) method for interpretation. The predictive performance of the models was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. Two rounds of reader studies were conducted (first round: independent reading; second round: guided reading) to validate the clinical application value of the DLR model, with results compared with FNA biopsy outcomes.

A total of 1561 radiomics features and 256 DL features were extracted from the original images. The DLR model, leveraging the ResNet152 neural network, could effectively differentiate between benign and malignant nodules in the context of HT. The AUC, accuracy, sensitivity, and specificity of the logistic regression (LR) model, based on these features, were 0.917 [95% confidence interval (CI): 0.838-0.988], 85.7%, 81.8%, and 86.6%, respectively, in the validation cohort and 0.827(95% CI: 0.777-0.876), 93.9%, 80.9%, and 96.3%, respectively, in the external test cohort. The SHAP summary plot illustrated how feature values influenced their impact on the model, whereas the SHAP force plot showed the integrated impact of features on individual responses. Gradient-weighted class activation mapping (Grad-CAM)-generated heatmaps from DL models visually highlighted high-risk areas in HT. The DLR model outperformed five junior-level physicians in terms of diagnostic accuracy, sensitivity, and specificity in the validation cohort. The diagnostic performance of all clinicians was significantly enhanced with the assistance of the DLR model, with no statistical difference detected compared with the FNA biopsy results.

The DLR model combined with the SHAP and Grad-CAM method can improve the diagnostic performance of radiologists in identifying benign and malignant TNs in the context of HT. The diagnostic efficacy of this visualization model is comparable to that of FNA cytology combined with gene mutation testing. The DLR model can enhance the diagnostic ability of radiologists in differentiating between benign and malignant TNs in the context of HT, thereby minimizing unnecessary biopsies. Additionally, it can aid clinicians in making personalized decisions regarding the necessity of biopsy or even surgery by providing intuitive visual explanations.