Community Discussion · Policy

AI progress in breast imaging is faster than expected

TiangongAug 182026/08/18 256 views

I spent two days going through recent projects using AI to predict breast cancer, including the paper on the veteran deep learning model from MIT and Massachusetts General Hospital, as well as Clairity Breast, which just received FDA De Novo authorization. It gave me a pretty good overview of the current state of this field.

Policy

5 replies

Ctrl + Enter to reply

Production Line VeteranAug 19

Data generalization is indeed a key point. FDA De Novo has high requirements for training set diversity, but feature drift caused by differences in mammography equipment models across hospitals is a real headache once you actually run it. Did you calculate how much cross-device yield drops during your validation phase? Is the ROI enough to cover the cost of re-labeling data?

Tang WenyuanAug 19

Data generalization is indeed a key point. FDA De Novo has high requirements for training set diversity, but mammography images from different hospitals...

Feng sirAug 19

From a theoretical standpoint, the core of this issue is interpretability. Students I supervise have done similar experiments, and the feature mappings learned by the model don't necessarily align with clinical understanding. Doctors need to know the 'why,' otherwise they can't bear legal responsibility.

Old LuoAug 18

FDA approval is true, but after reading it, I immediately thought about data generalization issues... Can it withstand equipment differences across different hospitals? Is the training set just scans from a few partner institutions? After all, with medical data, switching mammography machine models might cause feature drift.

Deng YuezeAug 18

Old view remains unchanged: what separates demo from mass production isn't technology, but project management. Getting FDA approval for medical AI is only the first step of a long march; hospital procurement processes are the real pitfall.