Putting eye screening AI into your laptop, how does Healgoo Interactive accomplish it step by step?

On November 29, Leiphone.com made an exclusive interview with Healgoo Interactive for Healgoo AI products. The following is the published interview content.

There are not many companies that use AI to screen for fundus diseases. Guangzhou Healgoo Interactive Medical Technology Co., Ltd. (hereinafter referred to as Healgoo Interactive) may not be the most dazzling one. But it has gone a long way on this road.

Healgoo Interactive AI Eye Diagnostic System - Healgoo AI has been applied to the screening of fundus diseases at the grassroots level, providing practical help to patients.

On July 21, Healgoo Interactive conducted a free eye screening campaign with Healgoo AI products in Songhe nursing home in Haizhu District, Guangzhou. As long as a single fundus camera is used to photograph the left and right eyes separately, the data will be automatically transmitted to the Healgoo AI ophthalmologic diagnosis system for analysis, and a written eye disease diagnosis report will be quickly generated for reference by the ophthalmologist. The entire process is only required a few minutes.

Healgoo Interactive CTO Wei Meng introduced that Healgoo Interactive was established in September 2013. At that time, they had already discovered the huge market for fundus screening, and developed a telemedicine cloud platform called Eye Grader.

After screening sites and doctors at primary hospitals take good pictures of the fundus, they can upload photos to the cloud through the Eye Grader platform and hand them to doctors in the top three hospitals for manual reading. Screened patients can also be referred to a professional eye hospital for treatment through the platform.

According to Wei Meng, there are more than 400 hospitals in the country currently using this platform, and the radiation scope covers all provinces except Tibet. Some hospitals in Hong Kong, India and Australia are also using the Eye Grader platform now. The DR (diabetic retinopathy) screening project held by China Lifeline Express Foundation, China National Defense Blind Foundation, Remote Vision Group, Zhongshan Ophthalmology Center, and Australian Ophthalmology Research Center is also worked on the Eye Grader platform.

Thanks to these early efforts, Healgoo Interactive accumulated a large amount of fundus photo data in the early stage, which laid the foundation for the subsequent turn to AI reading.

Wei Meng believes that data sets, algorithms, talents, and hardware are the four major elements of AI research, with data sets being the most important. He said: "As an AI application company, our main goal is not to invent new networks and algorithms. Everyone adopts more conventional algorithms, which are essentially similar in nature, so the main competition is the data set. Our advantage lies in having a large number of data."

Wei Meng said that Eye Grader is a manual reading platform. The fundus photographs on the platform were reviewed by two professional interpreters certified by UK NHS. As a result, Healgoo Interactive’s data annotation quality is very high and even exceeds that of Kaggle’s data set.

It is understood that Healgoo Interactive has divided the fundus photographs into five grades according to the severity of the disease, and the number has been evenly distributed. There are 200,000 images randomly selected for model training. In addition, other 70,000 images were also prepared for model testing.

In 2014, the Eye Grader platform was gradually expanding, and the limited readers could no longer cope with such a huge amount of reading workload. Healgoo Interactive began to try to use artificial intelligence technology to improve the reading efficiency.

At that time, the fire of deep learning has not yet ignited, and Healgoo Interactive chose to start researching on machine learning algorithms based on feature recognition. Feature recognition algorithm needs to identify the optic disc, optic cup, hemorrhage, exudation and other lesions, measure and calculate the relevant parameters and standards for comparison, and then define the disease's severity level according to the area and number of lesions.

After a period of exploration and research, Healgoo Interactive found that the robustness of this algorithm was not strong enough. For some fundus photos, its recognition is very good, but for other fundus photos, its recognition accuracy is difficult to reach the application level.

Wei Meng told Leiphone.com (Wechat: Lei Phone Network) that the feature recognition algorithm mainly has the following defects:

  First, the identification of lesions is not accurate enough. For example, it is easy to miss a bleeding point, or to mistake a part of a blood vessel as a bleeding point, resulting in a bias in the judgment of the severity of the disease.

  Second, the system has a "cask effect". Taking DR screening as an example, DR not only depends on the bleeding point but also observes the exudation. So two different models are needed here. Finally, the output of the two models is combined to draw conclusions. The accuracy of the conclusion often depends on the model with the worse performance.

  Third, the system has "dependency". For example, before calculating the area of the bleeding, the blood vessels must be extracted and removed. If there is deviation in the previous calculation step, it will inevitably affect the following step.

  Fourth, some diseases cannot be defined by strict mathematical criteria. In the case of cataracts, the presentation of cataracts on the fundus photographs is the degree of blurring of the photographs, which is difficult to define by mathematical models.

“After comprehensive consideration, we believe that the feature recognition algorithm has a very big limitation. Just when Kaggle's competition came out, we discussed with our R&D team at Stanford University and finally obtained the top 15 Leaderboard with deep learning algorithm. At the end of the competition, we also optimized the algorithm based on the data set of the Eye Grader platform combined with the training experience in the competition, and now our algorithm has been greatly improved, with an accuracy of 96.7% for DR recognition task.” Wei Meng said to Leiphone.com.

According to Wei Meng, most of the current AI companies that develop fundus screening products still use feature recognition algorithms, but some of them use traditional machine learning algorithms, and some of them are feature-based deep learning algorithms. The feature recognition algorithm essentially divides the fundus photo into many parts and then looks for the corresponding feature values to compare with the standard. The convolutional neural network takes the fundus photo as a whole.

He said: "We don't need to tell the algorithm which feature points are in the image. Just let it know if each image corresponds to a negative or positive result. Then it can learn specific rules by learning a lot of data."

But how does the algorithm work out in the “black box”? Healgoo Interactive is also very curious. Therefore, they allow the algorithm to analyze the images and mark areas where it believes they have an effect on the diagnosis, and let it draw a heat map. The heat map shows that the algorithm is based on the same criteria as the ophthalmologist.

In addition, Healgoo Interactive also invited a variety of senior residents, attending physicians and chief physicians to conduct tests in the same data set as artificial intelligence used. The results showed that Healgoo Interactive’s fundus screening AI has exceeded the attending physicians in terms of DR recognition and is close to the level of chief physician; it even surpassed chief physicians in glaucoma and age-related macular degeneration. Wei Meng told Leiphone.com, that based on the test results, Healgoo Interactive has submitted the paper to the international journal Ophthalmology and is currently waiting for the review.

Healgoo Interactive products are currently capable of screening for four fundus diseases of DR, glaucoma, age-related macular degeneration and cataracts, with an independent model for each disease. In addition, Healgoo Interactive has specially trained an image quality judgment model to ensure the quality of fundus photos.

Despite this, in Wei Meng’s view, the phrase “AI replaces a professional doctor” is still not very reliable. He pointed out that Healgoo Interactive must first address the initial diagnosis of retinal diseases under three scenarios:

  First, fundamental screening of grassroots communities and remote villages.

  Second, many grass-roots hospitals only have ENT, and have no professional ophthalmology. ENT doctors usually have limited knowledge of eye diseases. Artificial intelligence can help them conduct a preliminary diagnosis, to determine whether the patient needs to be referred to professional ophthalmology.

  Third, diabetes patients usually need to have course control in the endocrinology department. Artificial intelligence can help endocrinologists who do not have the ability to read fundus photos to preliminarily judge the fundus lesions of the patient.

In the actual application of products, Healgoo Interactive encountered some challenges. First, the network conditions in many communities and villages are poor, so cloud-based artificial intelligence screening products are not practical. Secondly, the hospital has a strong sense of protecting the privacy of patient data and has concerns about uploading fundus photos to the cloud.

As a result, Healgoo Interactive spent a lot of effort on optimizing algorithms in addition to the cloud version of Healgoo AI, and developed an offline eye fundus screening system that can be run on laptops.

Wei Meng said: "The cloud version is easier to implement, and the cost is low. There is no need to purchase new equipment, you can use it after opening an account, and the calculation speed is faster. However, because of the poor network environment, the cloud version is not in the grassroots screening scene. Our offline system will be able to run on laptops with a market price of around CNY 5000. In the future, we will further optimize the algorithms, reduce their hardware requirements, and improve the system's economics and portability."

It is understood that Healgoo Interactive's offline fundus disease screening system can give results within 1-2 seconds after the photograph is taken, and in addition to ranking the disease, it also gives a corresponding confidence index. If the degree of confidence is relatively high, then the diagnosis can be basically confirmed; if the degree of confidence is only about 50%, the results are still open to question.

For enterprises, developing products is only the first step; how to find a mature and feasible business model, transform products into commodities and achieve profitability is the foundation of enterprise survival. This is the dilemma that many current AI medical companies are facing.

Lian Song, the market manager at Healgoo Interactive, frankly stated that the marketing of AI medical products is not easy, but Healgoo Interactive has unique advantages than other companies. It has covered the Eye Grader telemedicine cloud platform in more than 400 hospitals.

He said that Healgoo Interactive plans to integrate the Eye Grader platform and the fundus screening AI. In this way, it will be possible to cooperate with fundus camera manufacturers (Healgoo will provide additional value of AI to fundus camera manufacturers, and camera manufacturers will provide sales channels for Healgoo Interactive) to promote fundus screening AI to hospitals and grassroots screening sites. After the patient was initially screened through the AI grading, the fundus photographs were uploaded to the Eye Grader platform and diagnosed by the doctors of the major hospitals. The patients are referred within the medical association according to the severity of the disease.”

In the construction of medical associations, in addition to behavioral synergy, the interface of data is also very important. Due to the different systems adopted by various hospitals, equipment manufacturers are variegated, and the standardization and interconnection of data has always been a major problem in the construction of medical institutions. Lian Song said that in the future Healgoo Interactive can also provide a data acquisition system for hospitals at all levels. The system can interface with the ophthalmic examination equipment of various hospitals to achieve real data standardization and interconnection.

Lian Song believes that artificial intelligence products directly help medical institutions to manage chronic diseases, and their contribution to the medical industry is unlimited. It may be more valuable to use artificial intelligence to promote the construction of medical associations. This is also the goal that Healgoo Interactive is striving for.

Click to view: the source on Leiphone.com