Deep neural network vs. commercial algorithm in low-dose CT image reconstruction

nature.com

13 points·bookofjoe··7 comments

I would be worried that the deep neural networks would fill in the blanks with prior information from other patients, which might be incorrect. In other words, the neural network imagines what should be there, and does not faithfully represent what is there

Yes, this should be a major concern, especially since CTs are used to detect abnormalities. It's very likely that the vast majority of the training data will be what is considered 'normal' for any given part of the body.

Indeed, you need a known, understandable and explainable error mode and model.

ANN are not even close to that.

Bonus points if you describe how they react to random perturbation - how robust the results are.

It's behind a paywall but the use of the "commercial" in the title seems weird (the abstract mentions "commercial iterative" technique fwiw). It seems like any kind of algorithm can be commercial and making this part of the comparison is odd. You could have an open source iterative technique or a commercial neural network approach.

It's somewhat ironic also they make their code available on github but not the actual article.

https://github.com/hmshan/MAP-NN and the actual DNN looks like a somewhat shallow residual network.

Anyone wants to share the article?

sci-hub.tw does

Reply on news.ycombinator.com