XIX Spanish Congress of Pathology Spanish Club of

XIX Spanish Congress of Pathology Spanish Club of Applied Computing in Pathology Voice recognition in Pathology Marcial Garcia Rojo, M. D. , Ph. D. Pathology Dept. Complejo Hospitalario de Ciudad Real Barcelona, September 18 th, 1999

Multimedia in Pathology

New perspectives en LIS (LIS = Laboratory Information Systems) F Graphic Interface – Workstation: electronic signature, coding, cytology Q/C, technique requests, etc. – Clerical work: Registry, Archive, Consultations, etc. F F F Computer Aided Diagnosis Unique Clinic History “Outreach”: Primary care and remote labs

Artificial Intelligence & Cytometry F F F Decision Trees: tumoral markers (Decaestecker, 1997) Expert systems (semantic networks) breast carcinoma classification (van Diest, 1994) Artificial Neural Networks: Essential in cancer – – – Gynecological cytological Screening: PAPNET Dg/Pro Breast carcinoma (O'Leary, 1992; (Mat-Sakim, 1998) Dg. Gastric lesions benign/malignant (Karakitsos, 1998) Liver Displasia & Carcinoma (An, 1997) FNA breast (Cross, 1997), Thyroid (Karakitsos, 1996)

Main obstacles in Speech recognition F Linguistic Variability: Phonetic, Phonology, Syntax, Semantics, Discourse F User V. : Tempo, Pronunciation, Inflexion, Fatigue, stress, coarseness. . . F Channel Variability: Noise, transmission channel F Coarticulation: Phonemes context

Acoustic Models F Phonetic-Acoustic Model (phoneme units) Speaker independent systems. Too unreliable F Pattern-Template. Methods to build them: u Vector Quantization u Hidden Markov Models u Neural Networks

Language Models F Provides the knowledge source for the engine. It helps to predict the next word. Components: u LEXICON (VOCABULARY): Context (Professional) u GRAMMAR: Structure and format (Markov Models)

Training F Crucial to all systems: Speaker dependant/independent 1 Read Text - Vocabulary - Pronunciations 2 Neural Network Training (iterative process -5 - of speech) & Hidden Markov Model building 3 To develop grammar for phoneme sequences F Greater training, less word error rate (plateau)

Philips Speech www. speech. philips. com F F F F F No end-user solution (Cortex) Recognition Server Speech Magic 64. 000 wd(270 K) Synchronized edition/ playback First Pathology continuous system Accuracy >= 95% Speech. Mike (Micro&Trackball&Spk) Speech. Pad (software & hardware) Free. Speech 98 30. 000 words

F IBM Med. Speak for Pathology 25. 000 word vocabulary of pathology u Template customization u Integration (Active-X) with LIS; HL-7 interfaces u F F IBM Via. Voice 98 (Español) - Medical vocab. (38. 000) Voice allows total control. Vocab: 33. 900 ptas. 95% Accuracy. 30 minutes de training Dynamic Healthcare Technologies (Co. Path) & Talk Technologies (Talk. Station-Pathology™)

Lernout & Hauspie F L&H Kurzweil Clinical Reporter Medical vocabulary. Templates. Total control u Excellent knowledge base (Rosai Cancer Checklist) u License cost: $6000 - $8000 per user (1 million ptas. ) u F L&H Voice Xpress for Medicine Integration with Microsoft Word u Medicine and Pathology vocabulary. Text Macros. u Cost: $499 (80. 000 ptas. ) Pathology: $1499 (250. 000 ptas. ) u F F F Up to 640. 000 words vocabulary Accuracy > 95%. 140 words / minute Requires Pentium II 266

Dictaphone Corporation Enterprise Express Voice System Integrated Voice and Data Management System F Explorer (Speech. Magic de Philips & Cortex path. sys. ) Client Viewer, Job Lister, Reporter (SQLQ). . . HIS integration, Fax distribution, E-mail, etc. Boomerang™: Software to send voice files (IBM Viavoice) Pathology context - 15, 000 basic pathology words F >95% Accuracy after speaker adaptation F F

Dragon Systems F F They have no product for Pathology but others (Voice Automated & Articulate Systems) use Naturally Speaking (Español) Naturally Speaking Medical Suite Active medical Vocabulary up to 55. 000 (240. 000) u Vocabulary Builder: Customization of medical specific vocab. u Dragon Best. Match: uses expanded speech and vocabulary models to get greater accuracy. u Dragon Command Wizard (macros, scripting language. . . ) u F F 160 words/minute. 95 -98% accuracy Pentium 200 y 64 MB RAM

F F F Broca: Philips engine. Surgical Pathology and Autopsy context First LIS for Pathology system with continuous dictation system (The Gold Standard™) Accuracy 98 -99%. 13. 000 pathology words Text & Formatting Macros For example, saying "Begin Microscopic" will indicate dictation for the microscopic section of the report and informs the system that the information eventually needs to upload to the Microscopic field in the LIS database.

Macros F Text Macros (E 3 -> Multiple fragments. . ) F Command Macros (keyboard/mouse) F Program Macros: Executes DLL, objets. . . F Structured Macros (templates)

Borowitz, Stephen M Pediatric Research 1999; 45: 120 A. F Computerized dictation systems are practical and fast enough to replace human transcription. F Differences in the types of errors: Human transcriptionists make spelling errors whereas computerized transcription makes word substitutions F Don't be surprised if your programs first rendition of 'cystic fibrosis' is something similar to 'sixty-five roses. ’ (Kovesi, T)

Other solutions F Only 7, 5% pediatricians (Spooner, 1998) F Medical Center contract Transcription Services to provide medical transcription at a cost of $0. 9 to $0. 19 per line. Local telephone call to dictate and the transcribed copy is delivered by modem. F Wireless Computer System Dragon Naturally. Speaking Mobile F Voice activation of instruments (Margossian et al, 1998)

SINOPTIC REPORT in Pathology

Interesting References F http: //home. nycap. rr. com/voice/ Voice Recognition For Pathology Michael Riben MRIBEN 1@NYCAP. rr. com F Teplitz C, Cipriani M, Dicostanzo D, Sarlin J. Automated speech-recognition anatomic pathology (ASAP) reporting. Sem Diagn Pathol 1994; 11: 245 -252 F Leslie, KO, Rosai J. Standarization of the Surgical pathology report: formats, templates, and synoptic reports. F Sem Diagn Pathol 1994; 11: 253 -257 Tischler AS, Martin MR. Generation of surgical pathology reports using a 5, 000 word speech recognizer. Am J Clin Pathol 1989; 92 (Suppl 1): S 44 -S 47. F Borowitz, SM. Computer-Based Speech Recognition as a Replacement for Medical Transcription. Pediatric Research 1999; 45: 120 A. F Hogan, R. Dragon Naturally. Speaking. JAMA Volume 280(15) 21 October 1998 pp 1369 -1370