Futurology

Accuracy in a classification context is defined as (N correct classifications / total classifications). So classifying everything as cancer would, in a balanced dataset, give you ~50% accuracy.

This article is indeed badly written PR fluff. I linked the paper in a sister comment. Both the confusion matrix and the ROC curve look phenomenal. Train/test/validation split seems fine too, as do the training diagnostics, so I'm optimistic that it isn't a case of overfitting.

Ofcourse 3rd party replication would be welcome, and I can't speak to the medical relevanve of the dataset. But the computer vision side of things seems well executed.

with an impressive 99.26% accuracy.

I feel this would be a blatant lie if it included a bunch of false positives.

https://mander.xyz/comment/17810389

While keeping the FPR low, our model keeps the TPR high, showing that it can accurately find real cases while reducing false alarms.

I'm not educated enough to know what recall means in this context, but there's tables with percentages for it in the page. (Would love an explanation; I'm not sure what to search for to get the right definition.)