Pediatric Bone Age Challenge -- data available

http://rsnachallenges.cloudapp.net/competitions/4

Also the you can find here a summary of the approach used by the winners.
https://stanfordmedicine.app.box.com/s/vhq1zop1867gr9rwnan4byj8lfxue173

8 Likes

Datasets are here

The winning solution is detailed in this blog post.

Terms and conditions here: http://rsnachallenges.cloudapp.net/competitions/4#learn_the_details-terms_and_conditions

Entrants retain all rights to algorithms and associated intellectual property they develop in the course of participating in the challenge.

Pretty straightforward competition, probably lowest hanging fruit in radiology. Think eventually will need to migrate away from using assessments based on radiologists matching images to an old atlas, better to use the actual ages of patients called “normal” to create a new standard. Could even supplement with other wrist radiographs not explicitly for bone age.

That would require some papers + time + acceptance to supplant Greulich and Pyle.

1 Like

Wish I’d had more time to optimize my parameters - winning solution was pretty close to what I tried.

Also looks like averaging of multiple crops may have been helpful for the inference stage - next time. :slight_smile:

1 Like

BTW, 4th place solution did not apply deep learning at all :open_mouth:

That’s cool :smile:
Unfortunately, I realized this competition when it was just over :joy:
I’ve tried afterwards in a quite simple way, my MAE seems to be around 10 so far :sweat_smile:
I’ve learnt a lot from the winning solutions – yup, next time :wink:

Some visualizations of people’s results:

http://rsnachallenges.cloudapp.net:5006/rsna_interactive

I also participated in the challenge. It was a very fun problem with a large public dataset (12600 images). Unfortunately for me, the test dataset was very different (different ground truth label, single center data instead of multi-center) compared to the training/leaderboard dataset. I definitely worked hard to get a good fit on the training/leaderboard dataset but it costed me a lot on the generalizability on the test dataset so I got 5.0 MAD, a little bit under the top 10 results for the test phase; big drop for me compared to the leaderboard rank. The test labels were definitely less noisy than the training/leaderboard labels. We usually learn something important for every data science competition !

RSNA has the intention to continue to host an annual competition of this kind. I hope many of you will participate next year !