A fully autonomous robotic ultrasound system integrated with artificial intelligence-assisted thyroid nodule assessment showed good agreement with conventional ultrasound and avoided 75% of unnecessary on-site examinations, but missed 19% of fine-needle aspiration biopsy recommendations, according to a study published in npj Digital Medicine. 

Researchers evaluated a fully autonomous robotic ultrasound system (FARUSS) combined with Aitrox-USIP, a commercially available artificial intelligence (AI)-assisted assessment platform.  

From March 2024 to August 2025, 262 outpatients requiring thyroid ultrasound examinations at three institutions in China underwent conventional ultrasound followed by FARUSS. The researchers compared image quality, assessment consistency, examination and analysis time, management recommendations, and patient acceptance. Conventional ultrasound-based expert assessment served as the clinical comparator. 

FARUSS used a 6-degree-of-freedom robotic arm and deep learning-based image perception to perform standardized thyroid scanning and provide real-time guidance. A medical assistant performed the FARUSS examination, and the complete examination was recorded as continuous video for subsequent analysis.  

The AI-assisted Aitrox-USIP platform used convolutional neural networks to analyze thyroid ultrasound images, identify nodules and their features, and assign American College of Radiology Thyroid Imaging Reporting and Data System (ACR TI-RADS) categories. Two expert sonologists with more than 10 years of experience reviewed the imaging data and reached assessments by consensus. For patients with multiple nodules, the analysis focused on the nodule with the highest ACR TI-RADS category, or the largest nodule when two or more had the same category.  

Robotic ultrasound shows agreement 

FARUSS showed good agreement with conventional ultrasound for thyroid gland measurements, with intraclass correlation coefficients ranging from 0.757 to 0.798, although agreement for isthmus thickness was poor at 0.126. The median robotic examination time was 202 seconds compared with 147 seconds for conventional ultrasound.  

Expert sonologists identified thyroid nodules in 180 patients using FARUSS images, and the key nodule matched that identified on conventional ultrasound in 154, or 86%. Among matched nodules, measurements, ultrasound characteristics, and ACR TI-RADS categories showed good agreement with conventional ultrasound. While FARUSS detected suspicious nodules in 78% of patients, its automated detection function identified the same key nodule selected by expert sonologists in only 37% of patients with nodules.  

Adding Aitrox-USIP reduced the sonologists' median analysis time from 254 to 183 seconds. Median total FARUSS scanning and analysis time was 383 seconds with Aitrox-USIP vs 457 seconds without AI assistance, although both remained longer than the 242 seconds required for conventional ultrasound scanning and analysis. 

Aitrox-USIP alone showed modest agreement with conventional ultrasound-based expert assessment for ultrasound features and ACR TI-RADS categories. Sonologists modified the AI assessment for 41% of identified key nodules. After expert modification, agreement with conventional ultrasound-based assessment improved, with weighted kappa values ranging from 0.754 to 0.879 for ultrasound features and the ACR TI-RADS category, except margin, for which kappa was 0.654.  

Screening strategy reduces on-site examinations 

Conventional ultrasound-based assessment recommended no fine-needle aspiration biopsy (FNAB) for 203 patients, follow-up for 33, and FNAB for 26. FARUSS produced management recommendations consistent with conventional ultrasound in 235 of 262 patients, or 90%. FARUSS alone avoided 75% of unnecessary on-site ultrasound examinations but missed five of 26, or 19%, FNAB recommendations.  

Aitrox-USIP alone avoided 72% of unnecessary on-site examinations but missed nine of 26, or 35%, FNAB recommendations. With sonologist modification, the AI-assisted approach avoided 74% of unnecessary examinations and missed seven of 26, or 27%, FNAB recommendations.  

Error analysis identified inaccurate measurements, incorrect ACR TI-RADS categorization, and missed key nodules as the main causes of discordant management recommendations. All five patients with an FNAB recommendation whose key nodules were missed by FARUSS had low-quality robotic images.  

 In an exploratory simulation, the researchers therefore referred scans with poor image quality to conventional ultrasound. This image quality-based triage eliminated all five missed FNAB recommendations in the low-quality image subgroup. Depending on the assessment approach, 66% to 67% of unnecessary on-site examinations could still be avoided. However, the researchers characterized these results as exploratory and hypothesis-generating.  

Limitations and next steps 

The researchers noted several limitations. FARUSS could assess the thyroid gland but not cervical lymph nodes. Conventional ultrasound-based expert assessment, rather than cytology, histopathology, or long-term outcomes, served as the comparator, meaning the study could not determine the diagnostic accuracy of FARUSS or Aitrox-USIP for thyroid malignancy. Initial probe positioning also required manual assistance.  

In addition, Aitrox-USIP analyzed selected stored frames rather than the complete FARUSS video, limiting its ability to incorporate temporal information across the scanning sequence. The key-nodule strategy also evaluated one key nodule per patient; in patients with multiple nodules, the researchers noted that this may not capture secondary nodules that could affect follow-up planning or routine documentation and may limit nodule-level screening reliability. 

The researchers concluded that while additional validation is needed, the combined system may provide a framework for decentralized thyroid nodule screening under professional supervision. “Further technical refinement and prospective validation with cytologic, histopathologic, and long-term longitudinal outcomes will help clarify its clinical value and facilitate broader real-world implementation,” they wrote.  

Dong Han and Tian-Shui Yu disclosed employment with Cobot Technology Co., Ltd., and Yuan-Ming Zhang, Chuan He, and Dong Wang disclosed employment with Shanghai Aitrox Technology Co., Ltd. The remaining authors declared no competing interests.