
It’s hard to imagine but speech recognition algorithms to generate radiology reports have been around for almost thirty years. In August 1987 a group of radiologists led by AH Robbins reported their findings with a “word recognizer” in Radiology. Given the limits of memory and speed then, they were restricted to a 1,000-word lexicon. However, by switching between lexicons (e.g. gastrointestinal radiology, neuroradiology and mammography), they were able to employ the system 88% of the time, with recognition reliability greater than 95%.
Over the intervening period the power of computers has increased exponentially as predicted by Gordon Moore, the co-founder of Intel Corporation in 1965. Moore’s Law states that over the history of computers, the number of transistors in an integrated circuit doubles approximately every two years. It is perhaps no accident that the vast majority of today’s personal computers have central processing units manufactured by Intel, its name a portmanteau of Integrated Electronics.
The UK Radiological Congress was recently held in Manchester and one of the papers delivered was entitled “Voice recognition system report accuracy.” The authors – Nazleen Gowdh and Karen Duncan from Aberdeen, Scotland – assessed the frequency of speech recognition errors in diagnostic reports and the failure of radiologists to identify these errors. The overall error rate was 10% where none of these were considered “major” errors, while the “moderate” and “minor” errors occurred in lengthy and complicated reports. The advice to their fellow radiologists was simple: carefully scrutinize your report before signing it off.
The Dragon speech understanding system was first developed by the husband and wife team of James and Janet Baker in the early 1980s when the hardware was not powerful enough to solve the problem of word segmentation, and users were obliged to pronounce one word at a time, separated by a brief pause. In time, this problem was solved and now the world’s best-selling speech recognition software – avaiable on both Windows and Apple personal computers – is called Dragon NaturallySpeaking and is produced by Nuance Communications.
There’s an interesting corporate connection between speech recognition software and breast cancer diagnosis hardware and it involves Carl Icahn, a businessman who has a reputation as a ruthless “corporate raider.” In April 2013 he acquired a 9% stake in Nuance Communications and then seven months later he bought 13% of Hologic, the world leader in breast imaging products. Given Icahn’s activist past, he has no doubt kept up the pressure on the management of both these companies.
It would appear that this system would have a positive feedback on the reporter’s diction in many fields, where the diction might resemble some doctors’ handwriting!