Ultrasound (US) diagnosis is crucial for examining various organs, including the thyroid. However, current US thyroid examinations heavily rely on sonographer skill and radiologist expertise, leading to variability in image quality and diagnostic results. The process is also physically and cognitively demanding for clinicians. To address these limitations, this research develops a fully autonomous robotic ultrasound system. The goal is to create a system capable of performing high-quality thyroid scans without human intervention, potentially improving the efficiency and consistency of thyroid US examinations and contributing to earlier and more accurate diagnoses of thyroid nodules. This autonomy aims to address the shortcomings of current methods by minimizing operator-dependent variations and patient-specific factors that introduce inconsistency and unreliability in US diagnostic outcomes. The level of autonomy in medical robots is categorized from 0 (no autonomy) to 5 (full autonomy). Existing US robotic systems typically fall into levels 0-3, with limitations such as manual path planning or reliance on preoperative images. This research aims to achieve level 5 autonomy, performing the entire thyroid scanning procedure autonomously.

Previous research in robotic ultrasound has explored different levels of autonomy. Level 0 systems involve manual probe manipulation; level 1 systems use visual servoing to track image features and compensate for patient movement during remote operation. Level 2 systems automate US acquisition along a manually planned path, while level 3 systems can autonomously plan and execute scans but require operator supervision. Existing approaches often rely on global target information from preoperative images or external sensors, which is challenging due to individual anatomical variability and patient movement. Methods incorporating image feedback or visual servoing for motion compensation have been proposed but may fail if target features are lost. While image-guided methods for robotic sonography in thyroid volumetry exist, they often use only a limited number of views. The increased autonomy also raises concerns about patient safety. Therefore, a fully autonomous robotic system adapted to clinical practice remains challenging due to the complexities of perception, planning, and control, alongside patient safety considerations.

The researchers developed a fully autonomous robotic ultrasound system (FARUS) for thyroid scanning. The system comprises a six-degree-of-freedom UR3 manipulator with a linear US probe, a force/torque sensor, and a Kinect camera for 3D skeleton joint tracking. The robotic scanning procedure consists of four phases: thyroid searching (TS), in-plane scanning (IPS), out-of-plane scanning (OPS), and multi-view scanning (MVS). **Thyroid Searching (TS):** The system initially estimates the thyroid location using skeleton joint data. Since anatomical variations exist, reinforcement learning is used to fine-tune the probe's position until the thyroid is identified via a segmentation model. **Probe Orientation Optimization:** Bayesian optimization, using image entropy as a loss function, optimizes the probe orientation to improve image quality. **In-Plane Scanning (IPS):** The probe scans upwards and downwards along the thyroid lobe, recording nodule locations using a segmentation network. **Out-of-Plane Scanning (OPS):** The system performs out-of-plane scans of previously identified nodules, while avoiding collisions. The scan halts if significant patient movement is detected. **Deep Learning for Segmentation:** Two separate deep learning networks were used for gland and nodule segmentation. The thyroid lobe segmentation network uses a ResNet18 encoder and UNet decoder, trained on the SCUTG8K dataset. The nodule segmentation network uses ResNeXt 50 encoder and UNet decoder, employing a two-step approach: pre-training on a larger dataset (TN3K and SCUTNIOK) and fine-tuning on a smaller dataset (SCUTNIK) from the specific probe used in the system. A custom loss function combining IoU loss, feature loss (weighting isogenic parts of nodules), and distance loss (constraining false positives within the thyroid gland boundaries) was used to improve segmentation accuracy. Grad-CAM was employed to visualize the model's attention mechanism. **ACR TI-RADS Classification:** The system automatically analyzes key characteristics of detected nodules (echogenicity, composition, margin, shape, echogenic foci) to estimate ACR TI-RADS levels for physician review. **Ethical Considerations:** The study adhered to ethical guidelines, obtaining informed consent from all participants and implementing safety measures to prevent harm during the robotic scanning process. These measures included US gel allergy tests, use of a collaborative robot that stops upon collisions, limiting the robot's workspace, force control to prevent excessive pressure, and a chair with wheels for participant comfort and mobility.

The FARUS system demonstrated the ability to perform high-quality, fully autonomous thyroid scans comparable to those of experienced sonographers. The system successfully integrated thyroid searching, in-plane and out-of-plane scanning, image quality optimization, and nodule detection. The deep learning segmentation models achieved high accuracy in identifying thyroid lobes and nodules, even for nodules with diverse echogenic characteristics. Bayesian optimization effectively optimized probe orientation, improving image entropy and quality with minimal adjustments. The FARUS system also demonstrated the potential to automatically classify nodules using the ACR TI-RADS scoring system. In a comparative evaluation with a professional sonographer's assessments on 19 patients, FARUS showed substantial agreement in nodule scoring and classification (10 out of 24 nodules had identical scores). Disagreements were mostly limited to a difference of one score point, primarily due to differences in echogenicity or composition classification. Furthermore, a subjective evaluation showed that most participants felt safe and comfortable during the robotic scan. A comparison of image quality metrics (entropy, centering error, mean confidence, LRIS) between FARUS scans and those from five experienced sonographers showed similarities, indicating the robustness of the robotic system. However, the FARUS was less efficient than experienced sonographers in terms of total scan time. The study also identified some limitations in detecting small and low-contrast nodules. Some potential false positives were also identified, highlighting the need for further refinement of the detection algorithm.

The results demonstrate the feasibility and potential of fully autonomous robotic ultrasound for thyroid scanning. The system's ability to perform high-quality scans comparable to those of experienced clinicians, its potential for accurate nodule detection and classification, and its patient-centered design highlight its significant advancements in the field of medical imaging. The use of deep learning for segmentation and Bayesian optimization for probe orientation optimization showcases the power of AI in enhancing the accuracy and efficiency of US examinations. The system's autonomous operation also holds promise for increasing access to high-quality thyroid US in settings with limited access to skilled sonographers. The high agreement with expert sonographer assessments, though with some discrepancies primarily in echogenicity and composition, suggests that FARUS can effectively provide data for ACR TI-RADS classification, informing clinical decision-making. However, the limitations identified highlight the need for continued improvement in the system's detection capabilities for smaller and low-contrast nodules, emphasizing the need for a larger, more diverse dataset for model training.

This study presented a fully autonomous robotic ultrasound system (FARUS) for thyroid scanning, achieving level 5 autonomy. FARUS demonstrates comparable image quality to human sonographers and shows potential for accurate nodule detection and ACR TI-RADS classification. While limitations exist, particularly in detecting small and low-contrast nodules, this work represents a significant step towards automating thyroid ultrasound, potentially improving diagnostic efficiency, consistency, and accessibility. Future work should focus on addressing these limitations by expanding the dataset, refining the segmentation algorithms, and integrating advanced image processing techniques to improve nodule detection and classification accuracy.

The study's limitations include a relatively small sample size of patients for the clinical validation of the diagnostic performance of FARUS. The dataset used for training the nodule segmentation model might lack sufficient diversity in nodule size and echogenicity, limiting the generalizability of the findings. The system's performance in detecting small, low-contrast nodules requires improvement. The study also did not fully address the potential impact of ultrasound image artifacts on the system's accuracy. The study’s recruitment bias, limited to a specific region in China, might influence the generalizability of the findings.