Self-learning algorithms analyze medical imaging data | Hacker News Reader

Self-learning algorithms analyze medical imaging data

tum.de

14 pointsgmays4 comments

“We learned about u-nets and they r gud”

Sounds impressive. But please allow me to rant here:

> are more accurate

Can we please stop using “accurate” as an adjective for models, especially when reporting on studies?

There are many other more telling metrics. Let’s start with precision and recall, and diagnostic ability. “Accuracy” tells nothing about the quality of a model.

/rant

The paper uses Dice Coefficient for the evaluation of their model which is common in such semantic segmentation tasks. The article is written for lay audience hence the use of "accuracy".

    Dice Coefficient = 2 * the Area of Overlap divided by the total number of pixels in both images.

0. https://www.nature.com/articles/s41467-020-19449-7#Abs1 1. https://towardsdatascience.com/metrics-to-evaluate-your-sema...

Thanks for digging deeper. I saw that it was from an academic institution (TUM) so the lack of useful metrics in the release admittedly ticked me off a little. Dice Coeff sounds very similar to Jacard Indexing for image tasks.

Reply on news.ycombinator.com