May 16, 2018
Skychain: AI for Respiratory Diseases Recognition
Medicine has always benefited from the forefront of technology. Technology advances like computers, lasers, ultrasonic imaging, etc. have boosted medicine to extraordinary levels of achievement. Artificial Neural Networks (ANN) is currently the next promising area of interest. It is believed that neural networks will have extensive application to biomedical problems in the next few years. Already, it has been successfully applied to various areas of medicine, such as diagnostic systems, biochemical analysis, image analysis, and drug development.
Artificial neural networks (ANNs) or connectionist systems are computing systems vaguely inspired by the biological neural networks that constitute animal brains. Such systems “learn” (i.e. progressively improve performance on) tasks by considering examples, generally without task-specific programming. For example, in image recognition, they might learn to identify images that contain cats by analyzing example images that have been manually labeled as “cat” or “no cat” and using the results to identify cats in other images. They do this without any a priori knowledge about cats, e.g., that they have fur, tails, whiskers and cat-like faces. Instead, they evolve their own set of relevant characteristics from the learning material that they process. Wikipedia.org
The problem
We at Skychain aim to create the infrastructure which will help to make the preliminary diagnostics more accurate. This kind of diagnostics is mostly tied to image analysis — pathology (imaging technique in medicine which deals with the nature of disease (structural and functional changes in tissue).
To understand how it is important here are some statistics:
Doctors can miss the first stage of lung cancer in 70% of cases when examining the x-rays of lungs. And still this is not the only case. The same problem goes for stained breast cancer slides, for example.
Our aim is to reduce the amount of medical mistakes due to implementation of ecosystem for ANN development and training. This will help to analyze the medical images with high accuracy and speed.
According to the research made by Indiana University in 2013, physicians using an artificial intelligence framework that predicts future outcomes would have better patient outcomes while significantly lowering health care costs. Cost per unit of outcome was $189, versus $497 for treatment as usual.
Our team started working to create an artificial neural network for respiratory disease recognition when the ICO ended in April, 2018 and by now we have achieved some great results.
Why did we choose respiratory diseases? The fact is that according to Eurostat Diseases of the respiratory system accounted for 7.7 % of all deaths in the EU in 2014.
Source: Eurostat
The solution
When creating our neural network we used the one programmed by Stanford researchers as a basis. We have changed the neural network by adding additional layers and by using higher resolution images. To make training process faster we pre-processed the original data. In the end, after several days of training, we got results comparable to the State of the Art.
State of the art (sometimes cutting edge) refers to the highest level of general development, as of a device, technique, or scientific field achieved at a particular time. It also refers to such a level of development reached at any particular time as a result of the common methodologies employed at the time.
We used different preliminary trained neural networks to obtain intermediate results. The best results were obtained when using NASNet Large (one of the best neural networks for image recognition).
After a few epochs we found out that the accuracy of neural networks remained on the same level. We lowered the learning rate and this gave us an opportunity to improve our result.
Then we created heat maps of signs, according to which the neural network determines the presence of the disease.
The images below contain the original roentgenogram, a rectangle, which the human specialist used to localize the damaged zones and the last thing is the results of the neural network work shown in a form of a heat map. Thanks to this we can value the quality of the diagnosis and prevent the possibility of retraining.
To depict the results we have created ROC curves for each disease.
Here are the results for each disease detected by our neural network (the closer to 1 — the more accurate is the neural network):
Atelectasis: 0.7820
Cardiomegaly: 0.8195
Effusion: 0.8769
Infiltration: 0.7175
Mass: 0.7481
Nodule: 0.6779
Pneumonia: 0.7643
Pneumothorax: 0.9045
Consolidation: 0.7525
Edema: 0.8773
Emphysema: 0.8797
Fibrosis: 0.6987
Pleural_Thickening: 0.7905
Hernia: 0.7977
The future
In the near future we plan to create 7–10 neural networks that can diagnose cancer and other diseases by analyzing magnetic resonance (MRI), computed tomography (CT), and X-ray images; electrocardiograms (ECG); and tissue sample images. Our neural networks will be hosted on the Skychain system and available for practical use. Until we officially launch the Skychain project, they will be available for free, so that anyone can try and use them.
We understand that as data is the lifeblood of the digital economy and for companies looking to apply AI to any number of areas, access to data is going to be one of the biggest challenges for our team.
To train machine learning algorithms one needs massive and clean data sets, with minimum biases. One needs also to keep in mind data privacy issues when it comes to harvesting personal data, particularly in light of the General Data Protection Regulation.
We are working in this field too as this is one of the most challenging parts of our project. We clearly understand that Skychain needs partners and users. It also needs extensive coverage in the specialized media to boost Skychain brand awareness.
