Diagnostic performance of artificial intelligence in interpreting thyroid nodules on ultrasound images: a multicenter retrospective study.

Background: Thyroid nodules are commonly identified through ultrasound imaging, which plays a crucial role in the early detection of malignancy. The diagnostic accuracy, however, is significantly influenced by the expertise of radiologists, the quality of equipment, and image acquisition techniques. This variability underscores the critical need for computational tools that support diagnosis. Methods: This retrospective study evaluates an artificial intelligence (AI)-driven system for thyroid nodule assessment, integrating clinical practices from multiple prominent Thai medical centers. We included patients who underwent thyroid ultrasonography complemented by ultrasound-guided fine needle aspiration (FNA) between January 2015 and March 2021. Participants formed a consecutive series, enhancing the study's validity. A comparative analysis was conducted between the AI model's diagnostic performance and that of both an experienced radiologist and a third-year radiology resident, using a dataset of 600 ultrasound images from three distinguished Thai medical institutions, each verified with cytological findings. Results: The AI system demonstrated superior diagnostic performance, with an overall sensitivity of 80% [95% confidence interval (CI): 59.3-93.2%] and specificity of 71.4% (95% CI: 53.7-85.4%). At Siriraj Hospital, the AI achieved a sensitivity of 90.0% (95% CI: 55.5-99.8%), specificity of 100.0% (95% CI: 69.2-100%), positive prediction value (PPV) of 100.0%, negative prediction value (NPV) of 90.9%, and an overall accuracy of 95.0%, indicating the benefits of AI's extensive training across diverse datasets. The experienced radiologist's sensitivity was 40.0% (95% CI: 21.1-61.3%), while the specificity was 80.0% (95% CIs: 63.6-91.6%), respectively, showing that the AI significantly outperformed the radiologist in terms of sensitivity (P=0.043) while maintaining comparable specificity. The inter-observer variability analysis indicated a moderate agreement (K=0.53) between the radiologist and the resident, contrasting with fair agreement (K=0.37/0.33) when each was compared with the AI system. Notably, 95% CIs for these diagnostic indexes highlight the AI system's consistent performance across different settings. Conclusions: The findings advocate for the integration of AI into clinical settings to enhance the diagnostic accuracy of radiologists in assessing thyroid nodules. The AI system, designed as a supportive tool rather than a replacement, promises to revolutionize thyroid nodule diagnosis and management by providing a high level of diagnostic precision.

Explainable Automated TI-RADS Evaluation of Thyroid Nodules.

A thyroid nodule, a common abnormal growth within the thyroid gland, is often identified through ultrasound imaging of the neck. These growths may be solid- or fluid-filled, and their treatment is influenced by factors such as size and location. The Thyroid Imaging Reporting and Data System (TI-RADS) is a classification method that categorizes thyroid nodules into risk levels based on features such as size, echogenicity, margin, shape, and calcification. It guides clinicians in deciding whether a biopsy or other further evaluation is needed. Machine learning (ML) can complement TI-RADS classification, thereby improving the detection of malignant tumors. When combined with expert rules (TI-RADS) and explanations, ML models may uncover elements that TI-RADS misses, especially when TI-RADS training data are scarce. In this paper, we present an automated system for classifying thyroid nodules according to TI-RADS and assessing malignancy effectively. We use ResNet-101 and DenseNet-201 models to classify thyroid nodules according to TI-RADS and malignancy. By analyzing the models' last layer using the Grad-CAM algorithm, we demonstrate that these models can identify risk areas and detect nodule features relevant to the TI-RADS score. By integrating Grad-CAM results with feature probability calculations, we provide a precise heat map, visualizing specific features within the nodule and potentially assisting doctors in their assessments. Our experiments show that the utilization of ResNet-101 and DenseNet-201 models, in conjunction with Grad-CAM visualization analysis, improves TI-RADS classification accuracy by up to 10%. This enhancement, achieved through iterative analysis and re-training, underscores the potential of machine learning in advancing thyroid nodule diagnosis, offering a promising direction for further exploration and clinical application.

Improving GAN Learning Dynamics for Thyroid Nodule Segmentation.

Thyroid nodules are lesions requiring diagnosis and follow-up. Tools for detecting and segmenting nodules can help physicians with this diagnosis. Besides immediate diagnosis, automated tools can also enable tracking of the probability of malignancy over time. This paper demonstrates a new algorithm for segmenting thyroid nodules in ultrasound images. The algorithm combines traditional supervised semantic segmentation with unsupervised learning using GANs. The hybrid approach has the potential to upgrade the semantic segmentation model's performance, but GANs have the well-known problems of unstable learning and mode collapse. To stabilize the training of the GAN model, we introduce the concept of closed-loop control of the gain on the loss output of the discriminator. We find gain control leads to smoother generator training and avoids the mode collapse that typically occurs when the discriminator learns too quickly relative to the generator. We also find that the combination of the supervised and unsupervised learning styles encourages both low-level accuracy and high-level consistency. As a test of the concept of controlled hybrid supervised and unsupervised semantic segmentation, we introduce a new model named the StableSeg GAN. The model uses DeeplabV3+ as the generator, Resnet18 as the discriminator, and uses PID control to stabilize the GAN learning process. The performance of the new model in terms of IoU is better than DeeplabV3+, with mean IoU of 81.26% on a challenging test set. The results of our thyroid nodule segmentation experiments show that StableSeg GANs have flexibility to segment nodules more accurately than either comparable supervised segmentation models or uncontrolled GANs.

Initialization of active contours for segmentation of breast cancer via fusion of ultrasound, Doppler, and elasticity images.

Active contours (snakes) are an efficient method for segmentation of ultrasound (US) images of breast cancer. However, the method produces inaccurate results if the seeds are initialized improperly (far from the true boundaries and close to the false boundaries). Therefore, we propose a novel initialization method based on the fusion of a conventional US image with elasticity and Doppler images. The proposed fusion method (FM) has been tested against four state-of-the-art initialization methods on 90 ultrasound images from a database collected by the Thammasat University Hospital of Thailand. The ground truth was hand-drawn by three leading radiologists of the hospital. The reference methods are: center of divergence (CoD), force field segmentation (FFS), Poisson Inverse Gradient Vector Flow (PIG), and quasi-automated initialization (QAI). A variety of numerical tests proves the advantages of the FM. For the raw US images, the percentage of correctly initialized contours is: FM-94.2%, CoD-0%, FFS-0%, PIG-26.7%, QAI-42.2%. The percentage of correctly segmented tumors is FM-84.4%, CoD-0%, FFS-0%, PIG-16.67%, QAI-22.44%. For reduced field of view US images, the percentage of correctly initialized contours is: FM-94.2%, CoD-0%, FFS-0%, PIG-65.6%, QAI-67.8%. The correctly segmented tumors are FM-88.9%, CoD-0%, FFS-0%, PIG-48.9%, QAI-44.5%. The accuracy, in terms of the average Hausdorff distance, is respectively 2.29 pixels, 33.81, 34.71, 7.7, and 8.4, whereas in terms of the Jaccard index, it is 0.9, 0.18, 0.19, 0.63, and 0.48.

Automatic initialization for active contour model in breast cancer detection utilizing conventional ultrasound and Color Doppler.

Regular examination of breasts may prevent and help to cure because breast cancer is treatable when it is detected early. Therefore, a breast cancer screening modality being sensitivity and cost-effective like ultrasonic imaging modality (US), is strongly required. In addition, the combination of a conventional US and its adjunct, Color Doppler has been proved for decreasing the rate of false-positive in breast cancer diagnosis. Thus, combination of these imaging modalities in a breast cancer segmentation would provide some benefits as well. An effective method for feature segmentation, active contour model has been widely utilized for decades. A crucial stage that affects the performance of active contour model is the initialization. This paper proposes a novel method for an automatic initialization of active contour model designed specifically for US-based imaging modalities. The method estimates an initial contour by utilizing the fusion of conventional US and Color Doppler. Examples and comparisons with three state-of-the-art automatic initialization methods are demonstrated to present the advantages of the proposed method. The evaluation results show high accuracy of initialization as well as fast convergence to features of interest.