Soyi Kwon MD, Presenter: Nothing to Disclose

Myung Jin Chung MD, Abstract Co-Author: Consultant, Samsung Electronics Co Ltd Patent agreement, General Electric Company

Youngjoo Moon BS, Abstract Co-Author: Nothing to Disclose

Jeongmi Cho PhD, Abstract Co-Author: Employee, Samsung Electronics Co Ltd

Chiyoun Park, Abstract Co-Author: Employee, Samsung Electronics Co Ltd

Kyung Soo Lee MD, PhD, Abstract Co-Author: Nothing to Disclose

Min Jung Park, Abstract Co-Author: Nothing to Disclose

Namhoon Kim, Abstract Co-Author: Employee, Samsung Electronics Co Ltd

Jaewon Lee PhD, Abstract Co-Author: Employee, Samsung Electronics Co Ltd

Eunsun Kim BA, Abstract Co-Author: Nothing to Disclose

Automatic speech recognition (ASR) system is increasingly used as a radiological reporting system; however, recognition errors occur frequently and thus manual corrections are needed. The purpose of this study was to compare the reporting time and error rates in radiologic reports generated with automatic speech recognition (ASR) system with those with conventional dictation transcription (CDT) system.

We collected 120 voice files used for dictation-transcription process of real radiologic reporting from 12 speakers. Ten medical transcriptionists with various experiences participated in this study. They were divided into two groups according to their experiences (Group 1, with > 10 years; Group 2, ≤10 years) as medical transcriptionists. Then, they transcribed twice the same dictated reports with the following methods; 1) they transcribed voice reports by using CDT system and 2) they corrected by listening to the voice files preliminary text reports generated by ASR system. We measured time for completing transcription by using both methods. Then, we scrutinized final text reports made by both methods in order to find any residual errors. We also calculated time ratios (RTs, elapsed time ratio for transcription over for recording) and compared them.

File length of recorded voice files ranged from 30 to 191 (mean ± standard deviation, 64 ± 34.7) seconds. Elapsed time for CDT ranged from 40 to 306 (113 ± 52.3) seconds. The time for ASR and correction ranged from 29 to 237 (96 ± 49.5) seconds. Average RTs were 1.8 ± 0.3 in CDT and 1.5 ± 0.4 in ASR and correction, respectively (P < .01, ANOVA). However, in Group 1, RTs were not different between the two systems (P = .20; RTs, 1.6 ± 0.2 in CDT and 1.5 ± 0.3 in ASR systems), whereas the RTs were significantly shorter in ASR than CDT systems (P = .02, RTs, 1.9 ± 0.3 in CDT and 1.5 ± 0.3 in ASR systems) in Group 2. Residual errors were 3.9% (319/8254 words) and 3.7% (305/8254 words) with CDT and ASR systems, respectively (P = .65).

ASR-generated report and further correction by transcriptionist is faster and equally accurate as compared with CDT system in radiologic reporting and transcription. However, in expert transcriptionists, CDT system is still comparable to ASR system in terms of efficiency and accuracy.

Particularly in novices in radiologic transcription, ASR system appears to be more efficient as compared with CDT system.

Kwon, S, Chung, M, Moon, Y, Cho, J, Park, C, Lee, K, Park, M, Kim, N, Lee, J, Kim, E, Automatic Speech Recognition in Medical Reporting: Comparison with Dictation Transcription System.  Radiological Society of North America 2012 Scientific Assembly and Annual Meeting, November 25 - November 30, 2012 ,Chicago IL. http://archive.rsna.org/2012/12043383.html