[simondotau]

Image compression for mammograms seems like a thing ideally suited to a straightforward ML task — a model trained to classify sections of an image which are outside of the area of relevance (i.e. not the breast) and classify details of high importance so that detail can be retained where it's needed.

Especially handy that it wouldn't require a new file format. It only requires the encoder to support variable compression. (I know Photoshop supported this for JPEG many decades ago.)

[bick_nyers]

I might be missing the details here, but everything outside the breast should be pretty much black and will get compressed very efficiently (high signal, low noise, easy to predict).

Foveated compression sounds like a super cool idea I've never thought about before.

Before that, you would need to get radiologists onboard with the idea of something being "diagnostically lossless" (think visually lossless), which is currently a hard sell (some promising research does exist on this)!