#rsna-intracranial-aneurysm-detection

1 messages · Page 1 of 1 (latest)

scenic junco
#

Pole

stone marsh
#

Can Anyone guide me how to approach this competition?

finite fable
#

Hi, I am.new to kagglr had a question regarding the submission criteria. What does it mean that my submission has to have a runtime less than 11 hours? Does it mean that the entire 2500 inferences should be completed in less than that?

And also the no internet rule. If my code relies on some packages like nibabel or idk torch then how does the no internet rule work.

Thank you for your help

scenic junco
#
  1. Yes, all hidden samples have to be inferenced in available time. 2. Near submit you can find '...' (more options) > install dependecies for all installable packages you will need.
coral meadow
#

Hi I'm new to this competition can anybody show me one of your notebooks maybe not the best one, but I'll get an idea how to work on this

silk solar
#

I keep getting throw errors on submissions, I am really struggling to solve it

scenic junco
#

From the provided submission code add your solution sequentially.

finite fable
modest phoenix
#

Hey everyone! I'm forming a motivated 5-person team for the RSNA Intracranial Aneurysm Detection Kaggle Competition and looking for passionate individuals to join me. This is a fantastic opportunity to work on a high-impact medical AI challenge.

We're aiming for top solutions and a collaborative environment. I'm looking to fill the following 4 key roles:

  • 1x Radiologist (Medical Domain Expert): Essential for accurate data interpretation, clinical insights, and ground-truth validation.
  • 2x Computer Vision AI Engineers (Deep Learning & Model Development): For building, training, and optimizing state-of-the-art detection models.
  • 1x Academic Researcher (Scientific Analysis & Literature Review): To provide robust research insights, explore methodologies, and contribute to the scientific framing.

If you're eager to contribute your expertise, learn from a diverse team, and tackle this critical problem, DM me to discuss! Let's build something impactful together.

#Kaggle #RSNA #MedicalAI #ComputerVision #DeepLearning #TeamUp

silk solar
finite fable
silk solar
#

About 10-12 minutes

#

Only on the submission though no error when I commit and run on my end

finite fable
#

I got errors are 40-60 mins of running so for my code it was most likely the data causing the issue.

#

For your case my best bet would be to run the inference pipeline on like 100 random data from train set. Check how much vram your gou is consuming and check if your pipeline catches all the errors. There's 39 images (on train set) which do not stack properly. So test your inferences on those images aswell.

#

GPU*

#

And another way to circumvent any errors (if its not your code) then wrap your entire prediction pipeline with try and except. Don't forget to return a proper data frame. However, if its your code causing those errors then this idea won't give you good result.

silk solar
#

ill keep trying thank you for the advice

solemn holly
#

Where do you get compute? This dataset is too large to train on a Kaggle GPU.

verbal compass
crimson briar
#

Hey!

  • Model trained ✅
  • Working on local example ✅
  • Currently submitting 🏁

However my notebook has been running for like an hour or so.... Is this normal?

finite fable
crimson briar
#

Thanks!

crimson briar
#

I figured out that it took like 20 seconds to make the predictions for the first 3 test files they supply.

My problem here is that most of the processing happens for the parsing of files. Does anyone have an idea how to do this more efficiently? Do you guys use multiprocessing to pawn multiple threads to read your files? Is this even viable here?

crimson briar
#

How do you guys read files? Most of the submission time i spent processing and reading the files…. However just reading every file using pydicom take more time that is requred. If there are 2500 images. And there are as many as 1000 channel per image… i have no idea how to read this….

finite fable
crimson briar
# finite fable Use multiprocessing to load dicom files. My entire pipeline for the three test i...

I have a multiprocessing pipeline. However Let's say an image has 1000 channels and is 500x500. That means we need to store 250000000 integer/ float values. Considering each value is represented by a 32 bit number... We get aroud a 1 GB of data per image. Is it even possible to process the entire image all at once/sequentially or is the only option to process it slice by slice? I know that all images are not of that size, however I have spoted some that are. My current pipeline process images slice by slice and outputs a prediction per slice. However I am just interesting if your way of doing is sequential if that is even possible?

full timber
gray echo
#

Hey all, I am a Biotech graduate. I also have knowledge and some experience in ML and data science. if you are looking for a team member, kindly consider me, I can still be a source of Biology knowledge. I want to get my hands dirty in ML 🙂 DMs open

verbal compass
#

hey, if anyone is looking for a team , i need teammates. 18th on LB right now , I have good ideas and need someone experienced to quickly run through them and implement them. Aim for a top 10 finish!

deep topaz
#

anyone still working on this competition now? im a beginner to kaggle but have been working in data science field for a while looking for teammate to cooperate

solid panther
#

hi everybody! I'm a fifth year AI Engineering student with experience in computer vision. I'm new to this competition and to Kaggle. DM me if you're interested in being teammates 😊

merry merlin
#

hey guys i keep getting this error while trying to use monai, does anyone have any clue why? Is it a data issue?

RuntimeError: applying transform <monai.transforms.io.dictionary.LoadImaged object at 0x7cf03a735150>

humble vigil
oblique flax
#

guys is anyone working on new approach, I am still thinking to use strong augmentation and retrain the trained model,like finetuning'

glacial osprey
#

its speading like a virus