An Artificial Intelligence Model Based on ACR TI-RADS Characteristics for US Diagnosis of Thyroid Nodules.

Background US-based diagnosis of thyroid nodules is subjective and influenced by radiologists' experience levels. Purpose To develop an artificial intelligence model based on American College of Radiology Thyroid Imaging Reporting and Data System characteristics for diagnosing thyroid nodules and identifying nodule characteristics (hereafter, MTI-RADS) and to compare the performance of MTI-RADS, radiologists, and a model trained on benign and malignant status based on surgical histopathologic analysis (hereafter, MDiag). Materials and Methods In this retrospective study, 1588 surgically proven nodules from 636 consecutive patients (mean age, 49 years ± 14 [SD]; 485 women) were included. MTI-RADS and MDiag were trained on US images of 1345 nodules (January 2018 to December 2019). The performance of MTI-RADS was compared with that of MDiag and radiologists with different experience levels on the test data set (243 nodules, January 2019 to December 2019) with the DeLong method and McNemar test. Results The area under the receiver operating characteristic curve (AUC) and sensitivity of MTI-RADS were 0.91 and 83% (55 of 66 nodules), respectively, which were not significantly different from those of experienced radiologists (0.93 [P = .45] and 92% [61 of 66 nodules; P = .07]) and exceeded those of junior radiologists (0.78 [P < .001] and 70% [46 of 66 nodules; P = .04]). The specificity of MTI-RADS (87% [154 of 177 nodules]) was higher than that of both experienced and junior radiologists (80% [141 of 177 nodules; P = .02] and 75% [133 of 177 nodules; P = .001], respectively). The AUC of MTI-RADS was higher than that of MDiag (0.91 vs 0.84, respectively; P = .001). In the test set of 243 nodules, the consistency rates between MTI-RADS and the experienced group were higher than those between MTI-RADS and the junior group for composition (79% [n = 193] vs 73% [n = 178], respectively; P = .02), echogenicity (75% [n = 183] vs 68% [n = 166]; P = .04), shape (93% [n = 227] vs 88% [n = 215]; P = .04), and smooth or ill-defined margin (72% [n = 174] vs 63% [n = 152]; P = .002). Conclusion The area under the receiver operating characteristic curve (AUC) of an artificial intelligence model based on the American College of Radiology Thyroid Imaging Reporting and Data System (TI-RADS) was higher than that of a model trained on benign and malignant status based on surgical histopathologic analysis. The AUC and sensitivity of the model based on TI-RADS exceeded those of junior radiologists; the specificity of the model was higher than that of both experienced and junior radiologists. © RSNA, 2022.

Clinical knowledge embedded method based on multi-task learning for thyroid nodule classification with ultrasound images.

Objective. Thyroid nodules are common glandular abnormality that need to be diagnosed as benign or malignant to determine further treatments. Clinically, ultrasonography is the main diagnostic method, but it is highly subjective with severe variability. Recently, many deep-learning-based methods have been proposed to alleviate subjectivity and achieve good results yet, these methods often neglect important guidance from clinical knowledge. Our objective is to utilize such guidance for accurate and reliable thyroid nodule classification.Approach. In this study, a multi-task learning model embedded with clinical knowledge of ACR Thyroid Imaging, Reporting and Data System guideline is proposed. The clinical features defined in the guideline have strong correlations with malignancy and they were modeled as tasks alongside the pathological type. Multi-task learning was utilized to exploit the correlations to improve diagnostic performance. To alleviate the impact of noisy labels on clinical features, a loss-weighting strategy was proposed. Five-fold cross-validation was applied to an internal training set of size 4989, and an external test set of size 243 was used for evaluation.Main results. The proposed multi-task learning model achieved an average AUC of 0.901 and an ensemble AUC of 0.917 on the test set, which significantly outperformed the single-task baseline models.Significance. The results indicated that multi-task learning of clinical features can effectively classify thyroid nodules and reveal the possibility of using clinical indicators as auxiliary tasks to improve performance when diagnosing other diseases.

Decoupling convolution network for characterizing the metastatic lymph nodes of breast cancer patients.

PURPOSE: The dual-energy computed tomography (DECT) technique is an emerging imaging tool that can better characterize material features and has the potential to be a noninvasive means of predicting lymph node metastasis. The purpose of this study was to establish a DECT-specified quantitative approach based on a neural network to characterize the sentinel lymph node (SLN). METHODS: With IRB approval, we retrospectively collected a total of 229 patients (100/229 metastasis) with biopsy proven breast cancer in this study. The chest and axillary spectral CT examinations were performed prior to the axillary lymph node (ALN) surgery. A decoupling convolution network with 11 ROIs from sequential keV (40 to 140 keV with 10 keV increment) was proposed to explicitly extract the spectral and spatial features in a DECT to predict the lymph node status. Focal loss was introduced as the loss function. The metric of the slope of the spectral Hounsfield unit curve measured at the venous phase was used as the baseline approach in comparison to our approach. In additional, a logistic model with radiomic features was also compared to our approach. The area under ROC curve (AUC) was used as the figure of merit to evaluate the classification performance. RESULTS: By introducing spectral convolution and focal loss, AUC on test set could be improved by 0.15 and 0.01 separately. Compared to the slope of the spectral curve with the average AUC of 0.611 and radiomic model with AUC of 0.825, the proposed approach demonstrates a considerably better performance, with test set AUC value of 0.837, by using decoupling spectral and spatial convolution together with focal loss function. CONCLUSIONS: We presented a new decoupling neural network based quantification method for DECT analysis, which might have potential as a noninvasive tool to predict metastasis lymph node status for breast cancer in clinical practice.