Ivan Zamorano

Acupunture Treatment

AI in Medicine | Medical Imaging Classification (TensorFlow Tutorial)

AI in Medicine |  Medical Imaging Classification (TensorFlow Tutorial)

this is what you’ve been waiting for hello world it’s Suraj and welcome to my new course on AI applied to business the most valuable course I’ve ever made we’re gonna build a simple AI app for doctors that lets them upload a picture of an eye and tells them whether or not that I has diabetic retinopathy if you’re looking to build a profitable AI startup or implement AI at your company in some way this is the course for you all my videos will be free and available right here on YouTube so make sure to subscribe to stay up to date with my latest content advances in medicine over the past few decades have improved health care immensely allowing doctors to more efficiently diagnose and treat diseases but doctors are still humans which means they understandably still make mistakes on the epic AF TV show house dr. House’s genius far exceeds his colleagues which implies that if all doctors were as smart as him diagnostic mysteries and unnecessary deaths in hospitals would drastically decrease but the reality is that the biggest difference between doctors is not their level of intelligence it’s how they approach patient problems and the type of health system that supports them this combination is what causes such wide variations in clinical outcomes and it’s the reason why machine learning is the best solution out there to improve doctors capabilities if you can’t beat them join them there’s so much potential here for example studies showed that over half of all women in the u.s. who get regular mammograms will receive at least one false positive which is a test that wrongly indicates the possibility of cancer in a 10-year period radiologists regularly disagree on their respective interpretations of medical images AI can do what no radiologist can it can learn from hundreds of thousands of medic images and is estimated to be up to ten percent more accurate than the average radiologist that accuracy gap will only increase as computing power gets cheaper and can be applied to any of the countless subfields of medicine not just radiology like what the startup vizhai is doing for brain scans doctors also have to interpret patient medical records which can be a very complex task NLP a branch of AI that helps computers understand and interpret human language can review thousands of medical records and output the optimal steps for evaluating and managing patients with many illnesses some doctors really are gifted at what they do like dr. oz just kidding an AI can learn from the best by watching them do their work if all doctors matched the performance of the top 20% patient deaths from a variety of diseases would decrease by the hundreds of thousands per year and wild doctors have natural biases AI is more likely to produce objective diagnosis for patients without preconceived socio-economic notions which can produce disparities in care ml will become an essential tool for doctors like the stethoscope has been and as more of the profession is automated human empathy and compassion for patients will become paramount to their success in the field also good looks let’s just be real so how do we pick a problem to solve the AI and healthcare market is expected to grow at an incredible forty eight percent compound annual growth rate in the next five years according to research and markets the tech giants like Google and Microsoft have massive amounts of data talent and computing power they have a huge advantage when it comes to building horizontal products these are products that can be applied to many use cases but as a start-up you can build a vertical product one that tackles a single problem very well since they don’t have the time to do that one way to do this is to find relevant problems in online communities where doctors congregate these can be forums slack channels subreddits Grey’s Anatomy chat rooms about McDreamy and see what kind of problems they’re having another way would be to call up or schedule a visit to a local practice to hear firsthand about the type of problems they’re having listening before acting is an important first step when finding the right problem to solve eventually we’ll come across a problem that is dealt with by multiple doctors let’s say for the case of this demo its diagnosing diabetic retinopathy correctly we can build a classifier to help solve this problem but first we’ve got to collect some quality data the foundation of all machine learning is having lots of quality data that’s how it learns very few companies actually generates and own medical grade data since collecting this data from patients is quite difficult so we’re gonna have to get a little creative we can search the web by searching public imaging datasets for diabetic retinopathy we’ll come across a few some will require us to apply and register before getting access so we can go ahead and do that and hope for the best alternatively if you’re a student at a university you can go find the nearest Imaging Research Center and ask the professor there they’re always willing to help with projects PubMed is also a great place to find biomedical research papers we can search for papers by the type of scan data we’re looking for and once we find a good one we can just email those researchers directly explain our project and how it could be mutual beneficial to both of us if we frame it more like we’re trying to help them then asking them for something they’ll be more likely to help us if we can get our hands on a quality data set ideally with labels because labelled data sets are always easier to learn from then we know that we have a shot at solving this problem but before we actually invest our time and energy into building an AI model we need to make sure that we’re able to get customers the easiest way to do this is to create a simple landing page we can find a template pretty easily online that asks for a simple email signup we’d explain our product in detail and the landing page and once we have it we can send it to potential clients in our case that would be doctors and medical companies we can find a directory of them online and email them one by one post it in facebook groups and see if we get any signups if we’re able to get a sufficient amount of signups then we know that there is indeed interest in our product and we can go ahead and start building there’s an entire ecosystem of libraries and services that help us build models our kyv sanity is a great tool to help a search for how the latest AI models have been applied to medical diagnosis in research labs once we find a cool paper we can use it as a guide to help build our own model it’s important to remember that there are so many tweaks and modifications we could potentially make to our model to improve it but the best thing to do is to first build a prototype that works as fast as possible and then iterate from there convolutional neural networks are a type of model that have proven to outperform all others when it comes to image classification which is our use case the fastest and easiest way to build a convolutional neural network is by using the chaos machine learning library it lets us build neural networks easily with its high level API each line of code corresponds to a different layer in our neural network we can search github for an image classification chaos model and use that as a base model notice how the parameters all seem like magic numbers how are we supposed to know which numbers are optimal this is the art of deep learning figuring out the right parameters for a certain type of model is the kind of thing that research papers focus on in this field sometimes a simple tweak gives rise to a breakthrough in accuracy we can use similar parameters to the paper we chose to start and train this model on our data set I have a macbook pro love you Apple let’s collab okay thanks bye and not a deep learning rig so I found that the best bet for me is to upload my model to Floyd hub which is a platform as a service that helps train deep learning models in the cloud we can do this in the command line with just a few commands what it’s doing is learning the mapping between the input data which are the medical images and the output data which are the diagnosis of the images once it’s learned this pattern given a new image it can accurately predict what the diagnosis is once it’s done training we can use Floyd Hough to create a basic API call where we merely send it an image and it will return the prediction of that image with any web framework of our choice we can build a single web page that asks the user to upload an image it’ll call the API and return the prediction or the user to see just like that from here we can make programmatic improvements add a better design and finally start collecting some sales hold on hold on hold on I’ve got three things for you to remember from this video hey I can help doctors better approach patient problems and improve the health system that supports them improving diagnosis interpreting patient medical history and drug discovery are just a few of the potential practices that AI can improve on and we can build a simple model using the Karos deep learning library then train and deploy it using floyd hub want to be friends hit the subscribe button and I’m all yours for now I’ve got to see my robot I mean doctor so thanks for

100 Replies to “AI in Medicine | Medical Imaging Classification (TensorFlow Tutorial)”

  • Thanks a lot for prepare this video series – My daughter just start her education in faculty of medicine in Thailand. I will bring this knowledge to help her for future career.

  • Interested in seeing how you ended up deploying the models. I ended up with a combination of ML engine | Cloud Functions | Firebase RDB Triggers which worked pretty well for my current MLaaS project >o

  • All the technologies are here, so each day we delay the development of open and free medical AI, thousands of lives are lost.

    So nice to watch this, right now working on app that would be able to diagnose lung cancer online.
    It would be free and decentralized web app.

    Would be nice to have WebVR dicom viewer, to better visualize images.

  • Siraj , Video is great and too much knowledgable.. , Can you please make Learning Path for Machine Learning.. ! ☺ From Beginner to Advance.. !

  • Been watching your videos for about a year Siraj. Love what you do, and completed my MSc in Data Science with your help 🙂
    Currently in my first Data Science job and have to say your videos helped me get started with understanding Deep Learning. Appreciate it bro!

  • The quality of the content you put out, Siraj got reeealy good. And, I love the episodes when you talk about using AI for business.
    Cheers bro, and keep up the great work!

  • Siraj, I'm a radiologist and prior computer science major. Really love your videos, and this one in particular! I'm taking Andrew Ng's deep learning course and hope to work in this space in the future! Look forwarding to checking out this course as well 🙂

  • I’m a radiologist and I’d like to get involved in AI. I need help with the programming. Let me know if you have experience and want to collaborate.

  • Dude. You rock my socks.

    "Some doctors really are gifted at what they do, like Dr. Oz. Just kidding."

    Made my day.

  • Pretty cool course man! Thanks for sharing this content 🙂

    Just a small request/feedback that I think it'll make the video cooler: Give more technical details about the project that you showed.

    I'd like to know more details about it.

    Besides, it's super cool to see health example and good uses for social data.

    All the best!

  • That's a great idea but having worked in the Healthcare IT field you should know you're treading on thin ice legally. An AI solution like this could get a Dr in A LOT of hot water if it made a mistake. And I don't mean to scare you but it could get YOU in hot water as well. Not with any criminal offense but you open yourself up to be sued, a Dr for malpractice and you for your part in it. Don't freak out. Add in some disclaimer and maybe talk to a lawyer.

  • Hi Siraj,
    I'm a high school student at the Loudoun Academy of Science, and I was interested in possibly creating models that would learn to recognize different insect bites and identify them (Dangerous bites vs ants vs mosquitos). As a high school student, do you know of any datasets I could access that could give me decent amounts of data to train the model with, or how I could find some? I was also considering using some sort of redundant analysis including some factors other than just images to help classify, such as the sensations(itchy, burning, stinging, throbbing) and an intensity attached. If you're able to help at all or have any ideas about this, I would really appreciate an email at [email protected]

    Thanks in advance for any reply and also for making this amazing course available for everyone!

  • What do you think about Neural Networks together with Genetic Algorithms? I'm trying to build a heart disease classifier by using a mix of both!

  • I am a doctor(MBBS, soon to be enroled in PG) , have learned python, now leaping towards ML. Thank you Siraj

  • Hi Guys, I am currently creating an application to help medical professionals detect diseases/features from medical images, through deep learning, and would like some insight from the medical community, as it would immensely help with the design of the application. Here's the proposal website: https://bit.ly/2KkqeJZ and here is the survey form: https://bit.ly/2r4CllJ

  • I actually love this guy. I'm about to graduate high school and have been unsure about my future. I'm interested in a lot of things in science, finance, and other fields but have always been into computer science, math, and statistics. Every time he makes a video I'm reassured I picked the right major of computer science. Plan to go into AI and machine learning eventually. Thank you Siraj

  • I am a doctor, currently further study in data science. I am very interested to know deep learning….tq bro !!!

  • Hey siraj great video! I was wondering how can we implement machine learning for classifying biomedical signals like ECG,EEG or BCG. Thanks in advance!

  • Hello Siraj. Thank you for all your entertaining and densely informative videos about AI. Can you do a playlist on AI and the legal domain or AI and Law. Can/will AI replace lawyers and judges? I think blockchain will be a huge part of this system. 🙂 Thanks in advance.

  • How do you verify that your training data is accuratet? How do you know you're not training your AI to misdiagnose?

  • The course literally started at 5:04 and ended at 10:07 ….Damn!!
    You teach more than a semester course in 5 minutes!

  • hi, guy. I am from China.I want to know how to begin with AI learning, I want to consult you, could you give me you contact methods,so i can commute to you?

  • Thank yo Siraj I have been watching your videos and learning so much. I made my own here using your course and lessons. https://biovision.ai

  • Guys I need your help. I really want to play with the diabetic retinopathy (DR) dataset but it seems I can't download it on kaggle. Does anyone else have the same problem? Where can I get those datasets?

  • I'm a radiologist working on AI/ML projects. I agree that AI/ML has a tremendous potential to HELP doctors including radiologists especially by removing the mundane repetitive work so they can focus more on quality rather than quantity. However, it also has an equal opportunity of being used for cost cutting for huge mega corps and insurance companies.

    The current trend is "AI better than Doctor/human X" which I believe is an improper metric designed to commercialize AI/ML. A better more accurate metric would be AI/ML helps doctor X save more lives. "AI/ML supervised by doctors or AI/ML supporting doctors in their practice" is the perfect set up but if we're not careful, it might turn into Machines Replacing Doctors everywhere because they're "cheaper for mass production, faster, don't complain, and are more readily replaced when obsolete." AI/ML is currently learning only with "correlation" and still cannot compete with the human brain's capacity for "cause-and-effect" which is a key differentiating factor between man and machine (but will businesses care about that?).

    I hope that more doctors become engaged in the #democratizeAI movement and your #100DaysOfMlCode.

    btw, love your videos Siraj. Keep learning a lot from them. I usually only watch and like but had to speak up about this since medicine and Radiology are near and dear to me. I really hope AI/ML is used for the betterment of health care, and not increased economic/production output.

  • I'm a BME Undergrad and I'm gonna be attempting to build a classifier this weekend at Shellhacks in Miami. If anybody's got medical datasets on anything please message me.

  • Mr Siraj Raval Great video. I am from Diabetic Hospital in India n a WHO collaborating Centre in India. Wld like to implement Google Tensor flow for Diabetic Retinopathy.
    Can u help us
    My email I'd is [email protected]

    Thanks n regds

  • Hello world… I have a question… Best way to display MHA files using python… Is? Display in the sense 3d visualising of MHA files..

  • Thanks man, an amazing tutorial. i'm doing my project on MRI brain classification using cnn. Datasets are in .mha file format. I'm using simpleITK library .But how can i generate images from .mha file?

  • AMAZING !!! I LOVE YOU. I'd love to make something for Image Processing Techniques used for Dental X-Ray

  • thanks a lot Mr. Siraj, I want to ask "How apply edge detection for medical image by Machine Learning especially with Genetic Algorithm"

    please, I will wait your answer with great enthusiasm

  • Preparing training data for radiology (medical) data: https://www.trainingdata.io ; https://www.trainingdata.io/blog/

Leave a Reply

Your email address will not be published. Required fields are marked *