Scholarlink

Background <p>Artificial intelligence-based image analysis has gained increasing attention in musculoskeletal ultrasonography. However, the reliability of general-purpose large language models (LLMs) for quantitative tendon assessment remains unclear. This study aimed to evaluate a ChatGPT-5-based interface for estimating patellar tendon cross-sectional area (CSA) on ultrasonographic images and to compare its outputs with clinician measurements.</p> Methods <p>40 adult patients with patellar tendinopathy were included. CSA was independently measured by two experienced clinicians in two sessions under blinded and non-blinded conditions. The same images were analyzed using ChatGPT-5. Reliability was assessed using intraclass correlation coefficients (ICC), with standard error of measurement and minimal detectable change. Agreement was evaluated using Bland–Altman analysis, mean absolute error, and mean relative error.</p> Results <p>Clinician measurements showed excellent intra-rater reliability (ICC = 0.989–0.997). In contrast, ChatGPT-5 demonstrated low internal reliability (ICC ≈ 0.33) and poor agreement with clinicians (ICC &lt; 0.05). CSA values were significantly underestimated by ChatGPT-5 (<i>p</i> &lt; 0.001), with a mean absolute error of 32.72 mm<sup>2</sup> and a mean relative error of 21.71%. Bland–Altman analysis revealed wide limits of agreement and a negative systematic bias.</p> Conclusion <p>ChatGPT-5 shows lower accuracy and repeatability than experienced clinicians in measuring patellar tendon CSA. These findings highlight the limitations of applying a general-purpose LLM to a task requiring precise anatomical delineation rather than reflecting AI performance in general. Accordingly, such models are not suitable for independent clinical use in quantitative tendon assessment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Feasibility and Reliability of ChatGPT-Based Estimation of Patellar Tendon CSA on Ultrasonography: Comparison with Expert Assessment

  • 期刊论文

  • 发表日期：

    2026年5月26日

查看全文
  • Ilhan Celil Ozbek,
  • Fatih Bağcıer

摘要

Background

Artificial intelligence-based image analysis has gained increasing attention in musculoskeletal ultrasonography. However, the reliability of general-purpose large language models (LLMs) for quantitative tendon assessment remains unclear. This study aimed to evaluate a ChatGPT-5-based interface for estimating patellar tendon cross-sectional area (CSA) on ultrasonographic images and to compare its outputs with clinician measurements.

Methods

40 adult patients with patellar tendinopathy were included. CSA was independently measured by two experienced clinicians in two sessions under blinded and non-blinded conditions. The same images were analyzed using ChatGPT-5. Reliability was assessed using intraclass correlation coefficients (ICC), with standard error of measurement and minimal detectable change. Agreement was evaluated using Bland–Altman analysis, mean absolute error, and mean relative error.

Results

Clinician measurements showed excellent intra-rater reliability (ICC = 0.989–0.997). In contrast, ChatGPT-5 demonstrated low internal reliability (ICC ≈ 0.33) and poor agreement with clinicians (ICC < 0.05). CSA values were significantly underestimated by ChatGPT-5 (p < 0.001), with a mean absolute error of 32.72 mm2 and a mean relative error of 21.71%. Bland–Altman analysis revealed wide limits of agreement and a negative systematic bias.

Conclusion

ChatGPT-5 shows lower accuracy and repeatability than experienced clinicians in measuring patellar tendon CSA. These findings highlight the limitations of applying a general-purpose LLM to a task requiring precise anatomical delineation rather than reflecting AI performance in general. Accordingly, such models are not suitable for independent clinical use in quantitative tendon assessment.