Scholarlink

Background <p>This study focuses on evaluating the effectiveness and reliability of GPT-4 in classifying radiological reports based on the Fazekas scale, a critical tool for assessing white matter signal abnormalities in brain MRI. We applied synthetic data creation and two specific GPT models, SinteticRMFazekasGPT and FazekasGPT, to generate and analyze 50 synthetic radiological reports. The study compared the performance of GPT-4 with the expert judgment of a neuroradiologist, for Fazekas classifications from brain MRI reports.</p> Results <p>Our analysis included contingency table and Cohen's Kappa for inter-rater agreement. The significance of the difference between the observed agreement and the expected agreement by chance was calculated, with a 5% threshold for a Type I error. The agreement between GPT-4 and the neuroradiologist was total (100%) regarding the Fazekas 0, with Fazekas 2 and with Fazekas 3. Out of the 15 reports with Fazekas 1, only 13 (86.7%) were correctly classified by GPT-4, while the remaining 2 (13.3%) were classified as Fazekas 2. Overall, the agreement was 96%, compared to an expected chance agreement of 28%. The Cohen’s Kappa value was 0.94 (<i>p</i> &lt; 0.001), indicating an almost perfect agreement.</p> Conclusions <p>We reported a novel application of GPT-4 to automatically obtain Fazekas classification from brain MRI reports. The results suggest GPT-4 as a promising supportive tool for obtaining Fazekas classification from brain MRI reports.</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

Automated Fazekas classification from brain MRI reports: an artificial intelligence approach with GPT-4

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月25日

查看全文
  • Carlo Augusto Mallio,
  • Andrea Carlomaria Sertorio,
  • Caterina Bernetti,
  • Federico Greco,
  • Gianfranco Di Gennaro,
  • Bruno Beomonte Zobel

摘要

Background

This study focuses on evaluating the effectiveness and reliability of GPT-4 in classifying radiological reports based on the Fazekas scale, a critical tool for assessing white matter signal abnormalities in brain MRI. We applied synthetic data creation and two specific GPT models, SinteticRMFazekasGPT and FazekasGPT, to generate and analyze 50 synthetic radiological reports. The study compared the performance of GPT-4 with the expert judgment of a neuroradiologist, for Fazekas classifications from brain MRI reports.

Results

Our analysis included contingency table and Cohen's Kappa for inter-rater agreement. The significance of the difference between the observed agreement and the expected agreement by chance was calculated, with a 5% threshold for a Type I error. The agreement between GPT-4 and the neuroradiologist was total (100%) regarding the Fazekas 0, with Fazekas 2 and with Fazekas 3. Out of the 15 reports with Fazekas 1, only 13 (86.7%) were correctly classified by GPT-4, while the remaining 2 (13.3%) were classified as Fazekas 2. Overall, the agreement was 96%, compared to an expected chance agreement of 28%. The Cohen’s Kappa value was 0.94 (p < 0.001), indicating an almost perfect agreement.

Conclusions

We reported a novel application of GPT-4 to automatically obtain Fazekas classification from brain MRI reports. The results suggest GPT-4 as a promising supportive tool for obtaining Fazekas classification from brain MRI reports.