Thoughts after taking deeplearning.ai’s AI In Medicine Specialization.
How Artificial Intelligence is empowering medical sciences
This is a three-course specialization. I will be talking about all of the three courses and my experience in learning about how AI is empowering medical sciences.
Introduction
Between a full-time job and working from home in these times, I spend my spare time learning about Convolutional Neural Networks and Computer Vision these days. At times, a great paper/video/course comes out that really intrigues me and I am instantly hooked. This course turned out to be that particular one.
A little background: I recently completed my deeplearning.ai specialization taught by
and loved how I got introduced to convolutional neural networks right from the basics. So, when deeplearning.ai recently announced it’s AI in Medicine specialization I was sure that this is gonna help me apply my learnings to a real-world problem.This course is taught by
, who is a Ph.D. candidate at Stanford University and his research work is in artificial intelligence (AI) technologies to tackle real-world problems in medicine.
“Working hard for something we don’t care about is called stress: Working hard for something we love is called passion.” -Simon Sinek
This specialization offers these three courses:
- AI for Medical Diagnosis
- AI for Medical Prognosis
- AI for Medical Treatment
Out of these, I’ll be talking in this article about my learning and views on the first course.
Course 1: AI for Medical Diagnosis
Imagine if you are given a chest X-ray image(unstructured data) and asked to develop a neural network to diagnose if a person has pneumonia or not?
OR
Given lab results(structured data) can you train a decision tree to estimate the risk of a heart attack?
Well, this course teaches you that.
Week 1: Disease detection with computer vision
We learn about classifying diseases from chest X-rays using a pre-trained deep neural network.
Key Concepts
- Data pre-processing: checking for data leakage
- Preprocess images properly for the train, validation and test sets
- Implement a weighted loss function to address class imbalance problem.
- Set up a pre-trained neural network to make disease predictions on chest x-rays.
Week 2: Evaluating deep learning models
As we can say that the decisions made by a neural net are high impact in medicine, it is important for us to identify when a model works on a patient and when it doesn’t.
If you don’t understand it you won’t be able to improve it
We can’t just account for accuracy as a metric to evaluate our model.
Consider this situation below:
This week dives more into mathematical concepts and terminologies which at once seem little intimidating but later you’ll get the hang of it.
I advise that one should write down the formulas taught in the video lectures and try and understand them analytically too. These formulas are gonna prove real handy in the upcoming assignments.
Week 3: Image segmentation on MRI images
Prerequisite: It’s recommended to have knowledge of Convolutional Neural Networks for this week’s assignment. It’s better to have an understanding of concepts like filters, padding, strides, and pooling layers.
In short, this week teaches you how to prepare 3D MRI data, implement an appropriate loss function for image segmentation, and apply a pre-trained U-net model to segment tumor regions in 3D brain MRI images.
Let’s dive deeper into the key concepts you will come across.
- You will perform image segmentation on 3D MRI data. You will learn about how to manage and represent 3D MRI data for computation purposes.
- You will take random sub-samples(or subsections) from a 3D image.
- Standardize an input image.
- Apply a pre-trained U-Net model.
- Learn about implementing a proper loss function for model training (soft dice loss).
- You will evaluate model performance by calculating sensitivity and specificity.
Course material and tools:
The lectures are delivered in a presentation format. I felt pretty comfortable in watching the lectures at 1.2x or 1.5x with concentration.
Each week is followed by quizzes and coding assignments in jupyter notebooks.
They have included in-between ungraded coding assignments in this course so that one can straightaway code the concepts learned rather than waiting for the complete week lectures to get over.
Assignments have a nice guided sequential structure and you are not required to write more than 2–3 lines of code in each section. Looking at the code structure helps in understanding how well-structured code is written and I encourage you to keep it a practice right from the beginning.
After the assignment is coded, it takes 1 button click to submit your code to the automated grading system which returns your score in a few minutes. There is a certain grade one needs to pass the assignment(not so difficult to get).
Jupyter notebooks are well designed and work without any issues. Instructions are precise.
P.S - I personally recommended taking the deeplearning.ai courses before jumping to this specialization.
