[jampekka]

It's quite easy to correctly classify 100% of benign cases as benign.

[cmcaleer]

If it's so easy, then why do people die from having lesions misdiagnosed as benign?

Even if the success rate of the human eye was in the 99.5%+ range, why not have an extra sanity check from an AI model?

[dijksterhuis]

> If it's so easy, then why do people die from having lesions misdiagnosed as benign?

You're confusing False Negatives with True Negatives. For Non-Benign (Positive) vs. Benign (Negative) classification:

* True Positive Rate (TPR): non-benign classified as non-benign.

* False Positive Rate (FPR): benign misclassified as non-benign.

* True Negative Rate (TNR): benign classified as benign.

* False Negative Rate (FNR): non-benign misclassified as benign.

> It's quite easy to correctly classify 100% of benign cases as benign.

You can engineer a 100% TNR if you just classify everything as the "benign" negative class. The FNR is going to be 100% too, but that doesn't matter -- you correctly classified 100% of benign cases as benign.

> why do people die from having lesions misdiagnosed as benign?

Because the FNR is not 0%. FNR is important. You probably want a decent TPR in there as well. And FPR can be very important too, depending on how life-changing/painful/invasive the treatment for a positive case is!

[OJFord]

Because this non-AI function 'correctly diagnoses 100% benign cases as benign':

    def is_benign(mole: Image) -> bool:
        return True

...but also misdiagnoses 100% malignant cases.

[potatoman22]

To contextualize, I think the tool in this article correctly diagnoses 97% benign cases as benign but misdiagnosis 22% of malignant cases.