I do a lot of MRI analysis including segmentation of small structures in the hippocampus called the hippocampal subfields. To collect these segmentations, we collect partial-field-of-view high in-plane resolution T2-weighted images on a 3 or 7 Tesla magnet. These sequences are generally only included in research protocols if the research specifically cares about hippocampal subfields...therefore they are rarely collected. There have been attempts to enable segmentation of these small structures using lower resolution T1-weighted scans, leveraging deep-learning or other models trained on concurrent T2w high resolution scans and the lower resolution scans within the same subject, allowing the model to predict the higher resolution information from the lower resolution inputs. This produces spectacularly beautiful segmentation on shitty data. Data whose resolution is about the same as the thickness of the structures you are segmenting or less. The problem is this: 1. The lower resolution image barely has any information in it on these smaller structures 2. The accuracy of the resulting segmentation depends entirely on how much the person fits the training distribution. But much research is on specific populations: children, autism, etc. 3. Some big names in imaging analysis tools have published these tools, lending their credibility to them. 4. The beautiful segmentations and (3) tend to convince non hippocampal experts that the resulting data is trustworthy, especially to an eager beaver researcher trying to maximize the impact of their already collected datasets.

I've rejected a number of papers for this.

But my point is this. Midjourney Medical might train a model to produce pretty images with this technique, but the more they need to depend on deep-learning models to get usable data, the more that the match between the training distribution and patient will matter.