We’re building Remene around a familiar problem: you receive your MRI images and a radiologist’s report, but making sense of them is difficult in many cases. Now, of course, many of us use ChatGPT, Claude, or other services like these to understand those terms and double-check the findings, just to feel a little more sure about what we’re reading.

One thing we discovered during our early tests is that this is not that easy to do with ChatGPT, Claude, or others. If you’ve already tried it with your own MRI, you’ve probably run into some of the same issues. A study can have hundreds of images, and some we tested had over 2.000. Getting all of those into a regular chat session is a problem. Even if you convert the DICOM files yourself, you can still run into upload and subscription limits. You keep trying, uploading smaller batches, asking it to fix something, and meanwhile you’re using up your subscription allowance before you’ve even got to the questions you wanted to ask.

You may have some luck with ChatGPT building a simple DICOM processor for the task. But these are scripts written for a one-off job, and getting something that runs is not quite the same as getting something you can trust with different MRI studies. From what we’ve seen, even though DICOM is a well-established standard, studies can come with all sorts of differences that these scripts don’t handle well. Sometimes you end up debugging the converter when all you wanted was to understand your scan.

But let’s say its superintelligence manages to build a perfect DICOM converter. There’s still the privacy question. DICOM files store lots of personal information: your name, date of birth, ID numbers, the clinic name, and more. It’s easy to think you’re just uploading files without realizing what else is inside them.

What we’re doing differently is putting a lot of effort into the DICOM processing pipeline. We de-identify the files (removing or replacing personal details), and keep the images and the metadata like which series they belong to, the view, whether fat suppression was used, etc. That way, we can reduce the personal information sent to the AI to the maximum. We keep sex information and age as a range, for example 30–39 rather than an exact age, because these can provide useful context for the analysis.

Secondly, we’re making sure the AI pipeline can get to the images it needs to analyze. We feed it images from the main MRI series, dozens and even hundreds, and give the agents tools to request additional images or information, or look at the same structure from another view.

And lastly, we also want this product to be a tool for understanding, but also for checking for potential misses or errors. We’re human, and sometimes it’s easy to miss a detail, especially when there are so many images in an MRI and so little time to go through them all. This doesn’t replace the radiologist or your doctor, and the AI can make mistakes too. It’s meant to be an additional check on your scan and report, with anything it flags being something you can discuss with your doctor.

The most critical point here is that this product is educational. It doesn’t give medical advice. It explains what the AI thinks it sees, what it can’t assess, and what might be worth discussing.

In the end, you’ve probably already spent some time staring at those images, reading the report again, and wondering what to ask. We hope Remene makes that part much easier, so you can go into the appointment with your questions ready.

Comments URL: https://news.ycombinator.com/item?id=50351830

Points: 8

# Comments: 0