search

juliandewit / kaggle_ndsb2017

Kaggle datascience bowl 2017
MIT License
622 stars 290 forks source link

some questions about evaluating the canditate nodules of the luna16_nodule_predictions #65

Open fulingly opened 4 years ago

fulingly commented 4 years ago

Hi,Julian, I use the model_luna16_full__fs_best.hd5 predict the nodules of the subset0,then I get the results in the luna16_nodule_predictions,and I start to evaluate the dection results,but I get a bad results,the sensitivity only is 32%, I don't know what went wrong.Can anyone give me some advice,thank you very much.

juliandewit commented 4 years ago

Hmm. Hi cannot say directly.. I have used model_luna16_full__fs_best.hd5 later and it works ok. Not that the smallest changes can give big problems..

fulingly commented 4 years ago

Hello, julian, thank you very much for the night gave me a reply, now I have put this experiment, but have been thinking, hope can do it, it is a great job, do you remember this model evaluating the sensitivity, perhaps can also, I reviewed methods of Luna16 evaluation standard of the game. Later, I selected some nodules with relatively large diameter on subset0 to evaluate, and the sensitivity result could reach about 80%. So I guess whether the detection effect is not good for nodules with small diameter? Just my guess, of course. I only used this model for the test, removed the model of unet segmentation of lung mass and other models,and directly used your model to test and evaluate on Luna 16, without retraining, which will have an impact on the results?Thank you again for your reply。

severous commented 4 years ago

@fulingly I don't konw if you sovled this problem. I also directly use this model to test and evaluate on luna16 without segmentation of lung mass and other models. thanks for your reply