A.I./Image Analysis Diagnose Eye, Lung Diseases

(Melmark, Pixabay)

23 February 2018. Researchers developed more efficient computational tools with artificial intelligence (A.I.) to diagnose retinal diseases and pneumonia by analyzing retinal scans and X-ray images. These new techniques, which generated results similar to those made by human experts, are described in yesterday’s issue of the journal Cell, with data and algorithms made available online to the research community.

The team from University of California in San Diego, with colleagues, from Texas, China, and Germany, are seeking more efficient and economical ways of harnessing the power of artificial intelligence, particularly machine learning, for diagnostics by analyzing medical images. While most machine learning techniques make use of readily available computing power and storage, applying this technology to diagnostics derived from images still requires vast numbers of images and laborious manual identification and classification processes. Not only are these current methods time consuming and expensive, but they also require a great deal of expert analysis.

The researchers led by UC San Diego ophthalmology and genetics professor Kang Zhang, are applying advances in machine learning to short-cut these processes. One of these advances is convolutional neural networks that apply filters to images, enabling learning algorithms to look for specific targets within the images. Each filter becomes a layer in the image, resulting in a layer-by-layer analysis, tied together with a map of the image’s key features. Convolutional neural networks make it possible to process images as pixels, with a classification of the image as output.

The second advance employed for this task is transfer learning, designed to extend machine learning to fields where data are more limited. Transfer learning applies lessons learned from other similar domains to the target fields, in effect giving the machine learning process a head start. The learning transferred with this technique still needs adjustments to make it applicable to the new data, but it enables the analysts to use a fraction of the data when starting with a blank slate.

Zhang and colleagues first analyzed optical coherence tomography, or OCT, scans to diagnose two leading causes of vision loss in the retina, macular degeneration and diabetic macular edema. The authors cite data showing nearly 10 million people in the U.S. have age-related macular degeneration, with 200,00 of those individuals developing a severe, blinding form of the disease known as choroidal neovascularization. And some 750,000 individuals age 40 and over have diabetic macular edema, a complication of diabetes, where fluid accumulates in the central retina.

OCT scans are a non-invasive technique that bounces light waves off tissue to construct detailed 2-D and 3-D images of the eye needed for accurate diagnosis of these disorders. The researchers trained their algorithms using about 207,000 OCT scans of people with macular degeneration and diabetic macular edema, as well as individuals with normal vision. The training routines taught the computer model  to recognize different parts of the eye’s anatomy rather than processing the entire eye at once, allowing for a faster and more efficient machine-learning process. The team then took a sample of 1,000 images from the database, using the same training routines, and found the smaller number of images could train the model with 93 percent of the accuracy as the full set of 207,000 images.

30 second diagnosis

In addition, the researchers added a separate occlusion test to better understand the factors contributing to the diagnostics. “Machine learning is often like a black box where we don’t know exactly what is happening,” says Zhang in a university statement. “With occlusion testing, the computer can tell us where it is looking in an image to arrive at a diagnosis, so we can figure out why the system got the result it did. This makes the system more transparent and increases our trust in the diagnosis.”

Once trained, the computer model can process images and make its diagnosis from the OCT scans and recommend a referral for treatment in about 30 seconds. The researchers tested its process with about 1,000 images from 633 patients, then asked 6 expert ophthalmologists not connected with the research team to evaluate the same images and make their diagnoses. The model’s diagnostics returns errors nearly 7 percent of the time, compared to about 5 percent for the human experts. In some cases, the model’s performance exceed that of some experts.

The team then applied these techniques to an entirely different medical issue, children with pneumonia, which according to data from World Health Organization cited by the authors, kills 2 million children a year under the age of 5 years. The researchers trained the computer model with more than 5,200 chest X-rays of children with viral or bacterial pneumonia, as well as healthy children. The team tested the model with images from 624 individuals, again divided between viral or bacterial pneumonia, and healthy lungs. The results show the computer model returns accurate assessments of the X-rays 93 percent of the time.

The researchers say the image analysis and machine learning techniques offer a platform that can be generalized to other medical images, such as MRIs and computed tomography or CT scans. To help make their techniques more widely available, the authors are making their algorithms and data available online.

More from Science and Enterprise:

• Brain Wave Data Harnessed for Open-Source Brain Model
• Microsoft Providing A.I. for Big Data Cancer Research
• Neuro Informatics Software Released to Open Source
• Super High-Resolution Photo Sensor Developed
• IBM, UC San Diego Partner on Healthy Aging

*     *     *