Radiology Reporting Software and Voice Recognition

What radiology reporting software actually covers

Reporting software is often discussed as though it means speech recognition. It does not. A reporting stack has four parts, and the vendors who sell each part have an interest in describing their own as the whole:

The capture layer turns speech into text. The template layer decides what a report of this type should contain. The structure layer holds the findings that other systems need to read as data rather than prose, such as laterality, measurements and assessment categories. The distribution layer gets the signed report to the referring office, the patient portal and the billing system.

Most conversations about reporting software are about the first layer. Most operational pain is in the fourth.

Why the capture layer gets all the attention

It is the visible one. A radiologist experiences the capture layer directly and continuously, so a bad one is intolerable and a good one feels transformative. That makes it the layer that gets demonstrated, benchmarked and argued about.

The other three layers are experienced by schedulers, technologists, referring offices and patients, none of whom are usually in the room when the software is chosen.

Recognition accuracy is not report accuracy

Speech engines are now good enough that the interesting failures are not mishearings of ordinary words. They are the failures that produce a fluent, grammatical, confidently wrong sentence.

That is not a hunch. A retrospective audit of 990 finalized reports from 19 radiologists was published in the Journal of Digital Imaging. It classified 482 errors: wrong-word substitution 37.6 percent, insertions 27.6 percent, deletions 14.3 percent, and outright nonsense phrases only 8.2 percent. Roughly four errors in five leave a readable sentence behind.

The same audit found error rates varied enormously by report type: 6 percent of computed radiography reports carried an error, against 38 percent of the rest. That is worth knowing before accepting one accuracy figure for a whole department.

What the pattern looks like in practice:

• a drug or device name transcribed as a similar-sounding common word

• a negation dropped, turning "no evidence of" into its opposite

• a measurement transcribed with the decimal in the wrong place

• an insertion that reads as a qualifier the radiologist never said

What the evidence does not support

Laterality is the error everybody names, and the one the data is least worried about. It is worth correcting, because it is repeated constantly.

A review of 2,923,094 reports (Luetmer and colleagues, Journal of the American College of Radiology) found 1,607 corrected laterality errors, 0.055 percent, of which 56 were major, 0.0019 percent. In that population they did not result in wrong-sided surgeries. More pointedly for this article: rates of major laterality discrepancies were unaffected by voice recognition software.

Voice recognition did change one thing. It was associated with a significant reduction in the time between a report being finalized and a correction being issued.

The honest reading: speech recognition changes the texture of errors, not the rate of the most catastrophic one. It produces more readable-but-wrong sentences, and it makes fixing them faster. Buying decisions should be made on correction burden and the quality of the review step, not on a fear of left-for-right that the evidence does not carry.

What actually reduces these errors

Three things, in order of effect. Structured fields for the values that matter most, so laterality and measurements are selected rather than spoken. Templates that force the report to answer the clinical question asked, so a missing element is visible as a gap. And a genuine read-back step, which costs seconds and is the only control that catches the fluent errors.

Where turnaround time is really lost

Report turnaround is usually measured from study complete to report signed, because that is the interval the reading system can see. The referring office measures something different: study complete to report in my hands. The difference between those two numbers is invisible to the reading system and is frequently the larger half.

We see this from the other side, because the report has to land in the RIS and then go somewhere. The delays are mundane and none of them are speech problems:

• the report is signed but the distribution job runs on a schedule rather than on signature

• the referring office's preferred delivery method fails silently and nobody is watching the failure queue

• the report is delivered but without the images or the priors that make it usable, so the office calls, which costs more time than the delay did

• an addendum is issued and the original is what was already distributed

Before evaluating a faster capture layer, measure your own signed-to-delivered interval. If it is a meaningful fraction of your total turnaround, a better speech engine will not move the number your referring offices care about.

How to evaluate reporting software honestly

1. 1

   Measure both turnaround intervals separately

   Study complete to signed, and signed to delivered. Do it before you talk to any vendor. Whichever is larger is the one your money should address, and it is frequently not the one you assumed.

2. 2

   Test with your own radiologists and your own templates

   A scripted demo uses the vendor's templates and the vendor's speech samples. Neither resembles your work. Ask for a trial with your report types and your accents, and measure correction time per study rather than accuracy percentage.

3. 3

   Decide what has to be structured, and insist on it

   List the values other systems need as data: laterality, assessment category, measurements, follow-up recommendations. Anything on that list which is captured as prose will be re-keyed by somebody, or lost.

4. 4

   Ask exactly how the report leaves the system

   Which interfaces, in which direction, triggered by what, monitored by whom, and what happens when a delivery fails. This is the question that separates a reporting product from a dictation product, and it is the one least often asked.