I've been meaning to write this post about ChexNet, a deep learning algorithm for a long time, mostly because hearing about it was the first time I thought-- "Wow, with skills in AI, I can achieve THAT." It was then that I realized how close AI could bring me to my goal of helping the underprivileged. The ease of the solution shocked me, and I realized that in the modern age, hundreds of solutions just like it had the potential to level out the differences between developed and developing countries.
Currently, to be diagnosed for a thoracic disease, an X-ray scan is taken, and a radiologist interprets the image. In countries with limited medical resources and personnel, patients are at a disadvantage and may not receive a diagnosis unless a doctor travels in or receives the scans electronically. Two-thirds of the world lack these resources, according to an estimate by the World Health Organization. Even where there is imaging equipment, medical experts are at a tremendous shortage. How can patients in these countries receive an accurate diagnosis without the hassle?
Here's how CheXNet works (or is it CheXNeXt?? Stanford's sites mention both).
CheXNeXT is an algorithm developed by students at Stanford that utilizes deep learning to detect pneumonia and 13 other physical pathologies. This 121-layer neural network can provide healthcare access to patients in developing nations and has taken the AI world by storm.
Pneumonia is discerned in chest X-rays as visible botches that are inflammation of the lung tissue due to infection. The algorithm learns to detect these patterns and differentiates them from other masses, such as tumors after training on a large number of images. To achieve this, the team built a convolutional neural network (CNN) to interpret the scan.
The Stanford team's (lead by Professor Andrew Ng) research paper highlights the mathematical function used in the algorithm and testing accuracy achieved. The most intriguing fact-- ChestX-ray14 (the highest-performing version) exceeds four practicing academic radiologists' diagnosis performance!
Reading this research paper was surprisingly easy (I guess my AI skills are ample now!) and motivating. Projects such as this inspire me to train my own neural networks to develop models for early detection & diagnosis of diseases such as skin lesions, breast cancer, or heart disease. It motivates me to learn more about AI because in its applications lie a world of possibilities for global change.