Manage dh compare opinions

app/manage-breast-screening/2026/03/comparing-image-reader-opinions/index.md

@@ -0,0 +1,126 @@
+---
+title: Comparing image reader opinions
+description: We've been looking at the steps we need second readers to take when reviewing breast screening images before they move on to the next case.
+date: 2026-03-27
+author: Danny Chadburn
+opengraphImage:
+  src: /manage-breast-screening/2026/03/comparing-image-reader-opinions/compare-recalls.png
+  alt: A 'Recall for assessment' first read shown next to a 'Normal' second read with buttons to keep or change opinion
+tags:
+  - beta
+  - prototype
+  - breast screening
+---
+
+We've been looking at the steps we need second readers to take when reviewing breast screening images before they move on to the next case.
+
+In breast screening, each set of mammograms taken is reviewed independently by two image readers. They each give their opinion on whether any further assessment is necessary. This process is usually conducted 'blind' so the reader who looks at the images second can't see what opinion the first reader gave. The other way of reading is 'open' where the second reader sees what the first said prior to giving their own opinion.
+
+There's no statistical evidence to prove that blind reading is a more effective way to identify breast cancer, but it's generally seen by breast screening offices (BSOs) as the preferred method so the second opinion given is unprejudiced by the first. 
+
+What happens next in the reading process largely depends on whether the opinions of the two readers match, or are different.
+
+## Laying out each scenario
+
+In our prototype, readers have four opinions they can give:
+
+1. Normal (no sign of cancer)
+2. Normal, but add details (when there's no signs of cancer but they want to add a comment)
+3. Technical recall (the participant needs to come back for more mammograms)
+4. Recall for assessment (a sign of cancer is detected that needs further investigation)
+
+They may also request images from prior mammograms or skip a case, but we'll be looking at those action in a separate task.
+
+We assessed the various opinion combinations and the relevant next steps. For example:
+
+* `IF` 1st read `IS` Normal `AND` 2nd read `IS` Normal `THEN` go to next case
+* `IF` 1st read `IS` Normal `AND` 2nd read `IS` Recall for assessment `THEN` review difference of opinion
+* `IF` 1st read `IS` Technical recall `AND` 2nd read `IS` Technical recall `THEN` review recall image requirements
+* etc, for 13 other scenarios
+
+Our proposed solution to handle these is to present a post-opinion screen after second reads that changes based on the combination.
+
+The core components of this are:
+
+1. A summary of the first read, including any associated details
+2. A reminder of the user's read
+3. An option to keep or change their original opinion
+
+We would also need to show readers the consequence of their decision (i.e. this case will go to arbitration).
+
+Our initial experiments with ways to display these elements included making choices around:
+
+* having an alert to show there's been a difference of opinion or stating this in the page heading
+* showing the read results next to each other or one above the other
+* asking a question about what to do next (with radio options) or presenting action buttons
+
+### A few examples
+
+![A yellow alert componet telling the second reader that their opinion was different to the first](compare-alert.png)
+
+![The opinions of the first and second reader presented on top of each other with an option form](compare-top-bottom.png)
+
+![Different opinions presented side-by-side with buttons to keep or change opinion](compare-buttons.png)
+
+## Our preferred design
+The side-by-side design with final opinion buttons will be tested with users from BSOs.
+
+![A 'Recall for assessment' first read shown next to a 'Normal' second read with buttons to keep or change opinion](compare-recall-normal.png "If a second reader says 'Normal' after the first has asked for a recall, they can keep their opinion but will be asked to add details to clarify why they do not agree. If they change to the first reader's opinion, they will be asked to add their own annotations to there is a full record of their read.")
+
+![Two 'Recall for assessment' reads with opinions per breast and related annotations](compare-recalls.png "If both readers ask for a recall, the second reader will be shown the first reader's opinion next to theirs. They may have spotted the same thing on the images, but have slightly different interpretations of what it is. If the second reader thinks they are talking about the same thing, they may choose to go with the first reader's annotations, or they may decide to keep their opinion and send all annotations through to the assessment stage.")
+
+![A 'Normal' first read and a 'Technical recall' second read with buttons to keep or change opinion](compare-normal-tech.png "Differences of opinion can also relate to technical recalls. In this instance, the second reader has asked for the participant to return so they can have one of the views retaken. However, if they see that the first reader was satisfied with the images and able to make a judgement, they may decide to go with their opinion. It's up to them if they believe the recall is necessary, but this could result in an reduction in unnecessary return visits for participants.")
+
+## Deciding when to show this
+
+As well as the design of this screen, we are also reviewing the best time to display it to users.
+
+### An early post-opinion screen
+
+We could present the first readers opinion to second readers immediately after they have given their opinion.
+
+The main benefit of this is saving duplicate effort. Users can 'adopt' the details that the first reader provided so they don't need to add the same information again if they've seen the same thing.
+
+However, this means only the general opinion is given blind, not the full details. We would need to discuss with clinical assurance and policy teams if this satisfies the requirements of guidance around double reading.
+
+There is also potential that users could 'game' the system. If they select an unlikely opinion, they could see the first read and then change their mind to match it. We can mitigate this by logging each time users switch opinions to identify any users doing this.
+
+### A late post-opinion screen
+
+The alternative option would be for users to give their opinion, fill in the relevant details (annotations, comments, etc) and then show the details submitted by the first reader.
+
+This would be a truer interpretation of blind reading. Readers would be required to give a complete read before seeing what their colleagues said which would mean a more comprehensive assessment.
+
+The major drawback would be a duplication of information. If the second reader identifies the same issues as the first reader, this would create multiple annotations which may then need a separate action to merge and reconcile.
+
+### A delayed post-opinion screen
+
+We also considered building a 'review list' that could be completed following a reading session.
+
+This would allow readers to stay in the 'reading flow' state where they can concentrate on reviewing images and giving opinions. This is unlikely to be developed further as we suspect it's better to interrupt this process and ask them to make an immediate decision rather than requiring them to re-familiarise themselves with cases later on.
+
+## The long-term benefits of change
+
+As well as seeing if this improves the image reading workflow for users, we want to find out if it can help address some of the other goals of the breast screening programme.
+
+### Resolving errors more efficienty
+
+As good as image readers are, they can't spot every sign of cancer. If we can tell them that a colleague has seen something of concern when they've given a normal opinion, it offers them an opportunity to take a second look.
+
+They may not be influenced by the first read and stick with their original opinion, but in instances where they've missed something obvious it can be immediately rectified without the need for further intervention.
+
+### Reducing the arbitration queue
+
+Different opinions that are reconsidered by the second reader will decrease the number that are sent for a third read (either by an individual or group).
+
+This may be a good or a bad thing. An extra opinion on a case is often valuable and we wouldn't want this new process to result in too many second readers changing their mind, just to avoid the arbitration step.
+
+### Removing duplicate effort
+
+The 'early' post-opinion screen described above allows users to adopt the details already added from the first read. While this may potentially save time, we need to understand if this is better overall than having the extra assurance of both readers identifying the same issue. 
+
+### Preparing for the future
+
+At some point, AI reads will become a thing.
+
+While it hasn't been specifically designed with AI in mind, our post-opinion review page would work well when cases have already had a first read by a robot. Users are presented with a first read which they can agree or disagree with, and may even be none-the wiser as to whether this was completed by a human or a machine.