Hello, In the training dataset and probably in the inference dataset, there are many exams with more than 8 images on a single exam (presumably a patient with breast prosthesis) that are not using the MLOID and CCID terminology for the included prosthesis views. Possibly some breast centers in the dataset didn't use that terminology. Can you confirm that presumption ? This potential non-conformity can be problematic for a comparison algorithm that need to compare an apple (MLO) with another apple (MLO) , instead of an apple (MLO) with an orange disguised as an apple (MLOID named MLO). Thank you, Alexandre.

Created by Alexandre Cadrin-Chênevert alexandrecc
You are right Christoph but according to my own practice in radiology, when there are 8 or more images from a screening exam, 95% of the time the patient has prosthesis. In the training dataset, about 50% of the exams with 8+ images have CC/MLO/CCID/MLOID and the other 50% just have CC and MLO images. That is a bit unusual statistically so that is why I was asking. But this could be related to a population selection bias.
For exams with more than the usual number of images, implant displaced views may be one reason. Another potential reason is that women have large breasts that cannot be completely captured in the detector. In these cases, the technologist will obtain multiple paneled views in CC and MLO projections to cover the entire surface area. In such a scenario, you could get 8+ views in a woman without implants. I'm assuming that that's what you're coming across and that the labeling (MLO vs MLOID) is correct.

Conformity of CCID, MLOID terminology in dataset for prosthesis exams ? page is loading…