Scholarlink

As radiologists around the globe contend with ever-growing workloads, the implementation of automated report generation presents a dual advantage: easing their burden and improving turnaround times. Empowered by Artificial Intelligence (AI)-driven speech recognition, our proposed method minimizes dictation errors, enhancing precision in report generation and thereby reducing the need for post-dictation corrections. In our research, we aim to pinpoint the most effective end-to-end Dictation-to-Report Pipeline. This entails assessing the performance of various combinations of NLP models (Facebook Word2vec Whisper, Facebook Word2letter, Kaldi, DeepSpeech), LLM models (Mistral-7B, Falcon-7B, Zephyr-7B, Qwen-14B), and Prompt Engineering for speech recognition and subsequent report generation. The achieved results include a Match Error Rate of 0.016, Word Error Rate of 0.131, Average Levenshtein Distance of 0.109, and Sentence Error Rate of 0.019. Our conclusions are further substantiated through a blinded qualitative evaluation by five radiologists, who assign validation scores to the generated radiology reports.

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

Streamlined Speech Recognition Model for Automated Radiology Reporting Employing Combined Automatic Speech Recognition Model, Large Language Model, and Prompt Engineering

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Manju Dabass,
  • Mohammed M. Jabeer,
  • Anuj Chandalia,
  • Dwarikanath Mahapatra

摘要

As radiologists around the globe contend with ever-growing workloads, the implementation of automated report generation presents a dual advantage: easing their burden and improving turnaround times. Empowered by Artificial Intelligence (AI)-driven speech recognition, our proposed method minimizes dictation errors, enhancing precision in report generation and thereby reducing the need for post-dictation corrections. In our research, we aim to pinpoint the most effective end-to-end Dictation-to-Report Pipeline. This entails assessing the performance of various combinations of NLP models (Facebook Word2vec Whisper, Facebook Word2letter, Kaldi, DeepSpeech), LLM models (Mistral-7B, Falcon-7B, Zephyr-7B, Qwen-14B), and Prompt Engineering for speech recognition and subsequent report generation. The achieved results include a Match Error Rate of 0.016, Word Error Rate of 0.131, Average Levenshtein Distance of 0.109, and Sentence Error Rate of 0.019. Our conclusions are further substantiated through a blinded qualitative evaluation by five radiologists, who assign validation scores to the generated radiology reports.