Python AI Organ Segmentation Tutorial
HTML-код
- Опубликовано: 4 авг 2023
- CHECK OUT MY NEW UDEMY COURSE, NOW 90% OFF WITH THIS CODE:
www.udemy.com/course/python-s...
Code:
github.com/lukepolson/youtube...
Relevant Libraries/Software:
MONAI: monai.io/
LIFEX: www.lifexsoft.org/
TCIA: www.cancerimagingarchive.net/
Get Your Billy T-Shirt: my-store-d2b84c.creator-sprin...
Discord: / discord
Instagram: / mrpsolver 
Наука
Wow. This video is almost down to the pennies precisely what I do at my cancer institute. Only I'm attempting to spot tumors with edge-detection nightmares working in 3D space with DICOM slices.
This is absolutely excellent. I'm going to send this video to every single person that asks me: "So what do you do for work?"
You using AI for your edge detection?
@@MrPSolver..can you believe we don't recieve comment replies,?? So hopefully a Like will be received 😅❤ Awesome work bro. You said you worked with Particle Acceleration??
Ohh, they don't have a reply option..
'some other key ' = 42. I am sold. Wonderful video
Wow, great, inspirational. I'm doing a bachelor's thesis on image segmentation on the thyroid. I wish to be as good as you one day.
Great video. Hoping you make more in this series. Very instructive. I noted you have 3D slicer on your windows task bar. Say I have a limited number of MRIs in which I want to segment the prostate (using this example since it came up in the video) as a whole gland as well as a specific subsection of the prostate representing a lesion. How can I use slicer to segment these images and then have these segmentations fed into python? Then how would I go about including the features of the segmented area in a model (along with, say, clinical features of the patient and disease) to predict some annotated clinical outcome like cancer recurrence?
I know this is not an easy question but if you're willing to make some more content on this, it would be awesome. Even start with some background information or pointing me towards some sources. Thanks!
This looks incredibly powerful, thank you for sharing. Do you have any advice for how to address the problem that the model/mask is missing parts of organs. Is it just retraining or do the tools also include the ability to manually curate and refine to rois (as an example, it looks like the brain missed a good chunk of one of the hemisphere?)
Sometimes as a postprocessing step you can fill in holes that exist within contiguous 3D regions. That would be under the assumption, I suppose, that there's no air pockets or something similar within the organ of interest I suppose.
Since Google is monitoring comments for Keywords... I'll say ,thanks for the A.i. content, I'm an aspiring Biomedical Engineer type guy.❤
Bro am doing research in radiology and I love you
There’s a lot of cool segmentation software out there! Recently I’ve been looking at PyRadiomics, maybe there is some interest here to do a video on it.
Is the offer for the UDEMY course going to be available again in the future?
Hello @MrPSolver, what if I want to only segment one organ not all?
I understand the loading the pre-existing model, do I need to remove the class labels from the model files (.json) that you talked about in the video??
How can we do that??
So helpful, thank you very much for your work!
Could you please link the video you uploaded that featured ‘hr lady’. I can’t find it anywhere now 😢
what if the resulting file later wants to be in the form of a dot. xml?
What application is it good to use?
Hello, @Mr P Solver, I would like to ask if it is possible to do a Python code with a train model who sings a certain song. Please
I feel like most people in the comments didn't try coding. I tried to replicate the same working environment but encountered countless conflicts. I know these things are a bit of luck but I don't think everyone is so lucky.
Sure, some luck but also a lot of troubleshooting with package changes, device specific things etc. Which error did you get?
Thanks for the great tutorial. Can this also be used for CT images of the chest area only?
Yup!
So helpful thank you ❤
Can this be used for histopathological images?
tried using my own ct scan dataset after getting his example to execute for me. "CT = image_loader(CT_folder)" wouldn't work with the following error "InvalidDicomError: File is missing DICOM File Meta Information header or the 'DICM'" . Weird. Maybe the ct scan in the hospital has different settings.
thank you
Thx
I think model name change or there are some issues in donwloading the model I mean wholebody_ct_segmentation downloading part
yes, i also have the same problem. did u find a solution ?
after some search based on what I found, MONAI label is no longer used for segmenting organs. they developed model called 'TotalSegmentor' that is used for this stuff right now. I'm not sure from my info so don't take it as real one if oy trying to segment organs too. I tried code in the notebook but faced problem when trying to download the model. error told that he can't find the http server. so, I tried to use it from 3d slicer itself but found the same problem when trying to use the server itself in MONAI label extension.
got an error from:
model.load_state_dict(torch.load(model_path))
found and used :
model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
that worked
Hey man I do rly wanna to get to this level in python I am a beginner and i am never consistent, so what I do?
To be good at anything you need to be consistent. The more you look into coding, the more tips and tricks you’ll find. It comes down to discipline, but also if u find it hard to be disciplined then maybe coding just isn’t for u. There are lots of other great skills to learn. Life’s short, don’t force urself to do the things u don’t enjoy!
Agreed with this. The reason I became consistent is because I program for a living as a PhD student.
U need mr shrek help
@@MrPSolver pls mek tutorial on how to be PhD stoodentist tyvm.
For me, the trick was to use python to make things that I needed for my job/studies. Necessity is the mother of invention. Pick some of the things that are the most irritatingly simple but still not automated tasks, those are ripe for python solutions.
16:20 why did you call it 42??😅❤
changed output_dir to mask folder and it didn't work even when I made the folder using forward slashes.
Used , after copying from datadir = 'C:\\Users\\trobe\\Pictures\\mri\\Data\\' #Data and using
"C:\\Users\\trobe\\Pictures\\mri\\Data\\mask" and it worked fine
Occasionally had trouble with this line, so I kept eval and just went looking for the folder in file explorer
based
31:57 - 100% cpu 99% memory full at 32GB - 7 minutes and counting for my 3 year old laptop versus his 1 minute . I have to find a smaller image
I used CMB-CRC-MSB-03400 - CAP 3.0 with its 218 images - executed in 6 minutes at 100% cpu and 95% memory at 32GB
I would like to propose help from experts in computer vision and image processing in general with regard to segmentation techniques applicable to real medical images.
Months ago, I uploaded a dataset of skull MRI images (on Kaggle) showing (separated by categories) tumor masses and was questioned about the images not having labels or markings.
Please understand that in a real scenario, professionals who need a diagnostic hypothesis have raw images in their hands, without markings, and precisely our effort must be to create tools that assist in diagnoses made through images.
So my appeal, if you know techniques for processing these images, any help is welcome, and more than that, tools developed today for the health sector can help save lives tomorrow...
My sincere thanks
No rap in the beginning? I'm out!
I run this and the code with it:
data = preprocessing({'image': CT_folder}) ...................
My computer hits 100% cpu /100% memory and then it stops running the code at 47 seconds. Intel coreI5 cpu 16GB ram with it max for this laptop. This laptop is from 2021.
Dell has laptops with photos of medical images on them - 64GB mem - need latest and greatest for this stuff.
Had memory of 16GB. Doubled it to 32GB and that portion of code worked. Got further along to prediction area and had to set the system to manage virtual memory because computer needed more than 32 GB of memory . I think it was looking for 64 GB. Trying prediction again.
*promo sm*