Development of reports preparation system for colonoscopy using speech-to-text technology

Traditional colonoscopy report preparation is time-intensive and prone to errors, relying heavily on memory and limited images. Enhancing the efficiency and accuracy of report generation is crucial to improving endoscopic workflows. This study aimed to develop and validate a Speech-to-Text (STT) system specifically for colonoscopy report preparation, enabling real-time documentation during procedures.

Data for system development included 12,691 colonoscopy reports and 190 min of acoustic data from real-world procedures. Additional training and validation used 1120 reports from four qualified endoscopists. The STT system employed deep neural network models for acoustic and language processing, incorporating medical words and noise-cancelation mechanisms. Word Error Rate (WER) and Word Accuracy (WAcc) were the primary outcomes for system evaluation.

During the experiment and modeling STT system, the developed system achieved a WAcc of 94.7% and demonstrated significant reductions in substitution, deletion, and insertion errors through iterative training and model adaptation. Validation with two endoscopists was conducted and achieved a WAcc > 95%. In addition, the endoscopists showed similar accuracies of 87.98 and 85.6%.

This study successfully demonstrated the feasibility of using an STT system for colonoscopy reporting, achieving high accuracy and potential for workflow enhancement. Future developments focusing on expanding data diversity and optimizing the system for broader clinical implementation are mandatory.