There are many players in the field of “AI lung disease diagnosis”, and the high false positive rate is questioned by the industry

【Image source: Digital Health】

Lung cancer is the most common cancer in China and worldwide.

According to industry reports, in 2018, there were 2.1 million new cases of lung cancer worldwide, accounting for 11.6% of all new cancer cases (ranked first); and 1.8 million deaths, accounting for 18.4% of all cancer deaths (ranked first).

Most lung cancer patients have already developed advanced cancer when they first visit a doctor, missing the best time for treatment. Therefore, early detection, early diagnosis, and early treatment are important ways to improve the cure rate of cancer.

Accurate screening of lung nodules is the first step in early diagnosis of cancer.

However, due to the small size, low contrast, and high heterogeneity of lung nodules, the diagnostic accuracy of ordinary chest X-rays is only 50%. Therefore, in recent years, chest CT scans have been widely used in nodule diagnosis, which can more accurately estimate the type of sarcoidosis, the degree of interstitial lung lesions, and lymphadenopathy.

For doctors, the process of reviewing CT images is time-consuming, labor-intensive and repetitive work.

A set of data shows that a tertiary hospital receives an average of about 200 patients with lung nodules every day, and each patient's examination will produce about 200-300 CT images, which means that radiologists need to read at least 40,000 images every day.

The pain points of the industry have brought huge market space for artificial intelligence technology. In recent years, using AI computer vision technology to identify lesions in medical images has become the business focus of many companies, many of which have launched lung disease diagnosis systems.

Leifeng.com noticed that there are many players in the field of lung nodule diagnosis, including industry giants such as Alibaba and Tencent, as well as AI medical imaging companies such as Linkedin Technology, Shenzhen Rui Medical, Tuma Shenwei, and Huiyi Huiying.

Linkron Technology: AI-Intelligent Diagnosis System for Pulmonary Nodules

LinkDoc Technology was founded in 2014. On August 3, 2017, LinkDoc AI-Pulmonary Nodule Intelligent Diagnosis System was officially launched.

The system uses a chest CT dataset annotated by radiologists working with LinkDoc. Based on deep learning and reinforcement learning techniques, it adopts a multi-channel, heterogeneous three-dimensional convolution fusion algorithm to improve the detection sensitivity and accuracy of lung nodules. It can ultimately automatically identify all nodules in CT images with a recognition rate of 91.5%, and calculate the medical characteristics of the nodules to assist doctors in writing diagnostic reports and aid in the final diagnosis.

It is reported that Link Technology has completed its Series D financing in the first half of 2018, with a financing amount of RMB 1 billion.

Yitu Medical: "Digital Lung" Intelligent Imaging Diagnostic System

Yitu Technology was founded in 2012 and completed a US$200 million Series C+ financing in June 2018.

Yitu Medical's intelligent lung cancer imaging diagnosis system, based on the Faster R-CNN algorithm and multiple optimization technologies, performs intelligent analysis of chest CT images, detects lesions, and describes the characteristics of lesions in multiple dimensions and makes differential diagnoses. It automatically generates structured reports based on clinical guidelines, which improves the efficiency and quality of doctors' film reading and enables patients to receive fast and high-quality medical services.

According to public reports, the actual clinical sensitivity of Yitu Medical's artificial intelligence technology has reached over 95%, among which small nodules less than 10 mm and tiny nodules less than 5 mm account for more than 40% of the detected nodules.

Shenrui Medical: Dr.WISE CAD medical imaging diagnostic system

Dr.WISE CAD medical imaging diagnostic system, a product of Shenzhen Rui Medical, has a sensitivity and specificity in detecting lung nodules, with an accuracy rate of 98.8%.

Shenrui Medical mainly develops medical application products based on artificial intelligence. It has currently launched intelligent medical imaging technologies and products for early screening and accurate diagnosis of benign and malignant diseases, providing intelligent solutions for the entire process of early cancer identification and detection, analysis and follow-up.

According to official sources, relevant technologies and products have been clinically tested in dozens of top-level tertiary hospitals and institutions, including Peking Union Medical College, Peking University Hospital, Chinese Academy of Medical Sciences Cancer Hospital, and Nanjing General Hospital.

On April 2, 2018, Shenrui Medical announced the completion of a RMB 150 million Series B financing led by Legend Capital and followed by Lenovo Star. So far, Shenrui Medical has completed three rounds of financing in just one year since its establishment, with the cumulative financing amount of Series A, Series A+ and Series B reaching RMB 300 million.

Tuma Shenwei: σ-Discover/Lung Nodule Intelligent Diagnosis System for Pulmonary Nodules

Tuma Shenwei was founded in 2015. In December 2017, it completed a RMB 200 million Series B financing round led by SoftBank China, with participation from Chende Capital and Delian Capital.

The company's σ-Discover/Lung Nodule intelligent lung nodule diagnosis system uses computer deep learning technology to analyze chest CT thin-layer scan images to help doctors quickly and accurately detect and identify lung nodules, make benign and malignant judgments, compare and analyze follow-up images, and automatically generate structured reports.

Currently, the product has been confirmed in many hospitals around the world, with a detection rate of over 97% and a compliance rate of over 90% for benign and malignant judgments.

Huiyi Huiying: AI intelligent lung nodule screening system

Huiyi Huiying AI intelligent lung nodule screening system compresses lung imaging diagnosis to seconds. It can automatically identify lung nodules in hundreds or thousands of frames of images, mark the size, location, density, and preliminarily distinguish between benign and malignant tumors and automatically generate structured imaging reports for doctors to review.

According to Leifeng.com, the system has assisted doctors in completing nearly 50 cases of lung nodule screening, with a sensitivity of over 95%.

Huiyi Huiying was founded in 2015 and is a medical imaging artificial intelligence diagnosis and treatment and big data service company. In October 2017, Huiyi Huiying announced the completion of a B round of financing of several hundred million yuan.

InferRead™ CT lung screening system

InferRead™ CT system from Infervision Technology can effectively explore the core features of lung cancer in the early screening of lung cancer, determine whether different sequence images have suspicious cancer features, and assist in the early diagnosis and treatment of lung cancer.

It is highly sensitive to easily missed nodules such as solid nodules and ground glass nodules below 6 mm, and can automatically identify and mark solid, calcified, partially solid, ground glass density nodules and other nodular lesions and provide comprehensive and accurate information, including nodule location, size, density, properties, etc. The system can automatically pre-process, click and view, automatically compare the before and after films, and view images in a linked manner.

In March 2018, Inspur completed a RMB 300 million Series C financing round led by Xianghe Capital, Advantech Capital, Yuansheng Capital, Sequoia Capital, and Qiming Venture Partners.

Vision Technology: Lung Sight CT Chest X-ray Intelligent Diagnosis System

Shijian Medical is headquartered in Shenzhen, and its product research and development includes three major directions: radiology imaging (DR/CT/MR) assisted diagnosis, radiotherapy target area assisted delineation, and pathology (cytology, histology) imaging assisted diagnosis.

In the field of radiology, the Lung Sight CT chest film intelligent diagnosis system launched by Vision Technology can detect lung nodules, automatically segment the lungs, locate nodules, calculate more than 100 imaging genomics information, classify nodules according to the Asian consensus, provide medical basis for doctors' diagnosis and follow-up nodule comparison, track nodule change trends, and predict whether nodules are benign or malignant.

In March 2018, Vision Technology completed its Series A financing led by Shenzhen Capital Group. In June of the same year, it completed additional round A financing with China Merchants Venture Capital as the investor. So far, Vision Technology has completed a total of RMB 100 million in Series A financing.

Ali Health: “Doctor You” AI system

On July 11, 2017, the medical AI "Doctor You" developed by Ali Health was officially released to the public at the Beijing Wanli Cloud Medical Imaging Center.

The "Doctor You" CT lung nodule intelligent detection engine was jointly created by Ali Health's algorithm engine team and Alibaba iDST's visual computing team. It combines medical knowledge and artificial intelligence technology to automatically identify and mark suspicious nodules, improve doctors' work efficiency, and reduce misdiagnosis and missed diagnosis rates.

According to the Ali Health official website, the "Doctor You" AI system includes a clinical medical research and diagnosis platform, a medical auxiliary detection engine, a physician capability training system, etc. Ali Health and Wanli Cloud demonstrated the real application scenario of the "Doctor You" CT lung nodule intelligent detection engine in the remote imaging diagnosis process.

Tencent Miying: Intelligent screening system for early lung cancer

Tencent Miying's team consists of more than 50 AI scientists with doctoral degrees or above, more than 400 AI application engineers, and several medical experts and product managers. The GPU computing cluster worth hundreds of millions of dollars supports the rapid iteration of deep learning networks. A single inspection of more than 500 images can be processed in 6 seconds.

It is reported that Tencent Miying has developed an intelligent screening system for early lung cancer using Tencent Youtu's advanced technologies in big data computing, image recognition and deep learning.

The 3D segmentation and reconstruction algorithm of the lungs in the preprocessing module of the system can process different source data generated by different CT imaging devices under different imaging parameter conditions. The multi-task 3D convolutional neural network algorithm can be used to process and identify lung CT images to achieve early lung nodule detection.

According to calculations, its sensitivity (recognition accuracy) for early lung cancer is over 85%, its specificity (recognition accuracy) for benign lung nodules is over 84%, and its detection rate for tiny nodules with a diameter greater than 3mm and less than 10mm is over 95%, which can help radiologists significantly improve the early cancer screening capabilities of lung CT.

LUNA evaluation: iFlytek and Ping An of China topped the list

The LUNA (LUng Nodule Analysis) evaluation is the most authoritative international evaluation in the field of lung nodule detection, and is also one of the most representative and most concerned evaluation tasks in the field of medical imaging.

In August 2017, iFLYTEK's artificial intelligence research achieved an average recall rate of 92.3% in the LUNA, an authoritative international medical imaging evaluation, and won the first place in the evaluation with a significant advantage and set a new world record.

After achieving world first place in intelligent speech, semantic understanding, machine translation and other fields, iFLYTEK has also become a world leader in the important field of combining computer vision and medical imaging.

In January 2018, the test results released by LUNA (Intelligent Reading of Lung Nodules) showed that Ping An Technology's intelligent reading technology under Ping An Group of China achieved a major breakthrough, breaking the world records for "lung nodule detection" and "false positive screening" with an accuracy of 95.1% and 96.8% respectively, both ranking first in the world.

This is the first time that Ping An Technology has won a world record in the field of medical imaging, and it is also a milestone achievement for Ping An of China in the field of artificial intelligence.

Ping An Technology's medical imaging team leveraged its comprehensive advantages in the field of medical diagnosis and creatively proposed a solution that deeply integrates robust control theory with artificial intelligence algorithms such as deep learning and transfer learning, ultimately achieving a breakthrough in the detection, positioning and identification of lung nodules. Currently, this technology is gradually being applied in actual medical auxiliary diagnosis scenarios in many tertiary hospitals.

New York University: Inception V3 algorithm identifies cancerous tissue

A study published in Nature Medicine in September 2018 showed that researchers used a large number of digitized pathology slide images to retrain Google's deep learning algorithm Inception V3. The AI's accuracy in identifying cancerous and normal tissues reached 99%, and its accuracy in distinguishing between adenocarcinoma and squamous cell carcinoma reached 97%.

In addition, the algorithm can identify the six common gene mutations in lung cancer from slice images, including EGFR mutations that already have targeted therapies, with an accuracy rate of 73%-86%.

Currently, it takes about 20 seconds to detect a digital slice on a single GPU. Researchers believe that with newer technology and multi-GPU parallel processing, the detection time for each slice can be reduced to just a few seconds.

Andrew Ng’s team: CheXNet detects pneumonia

Andrew Ng, a famous deep learning scholar, and his team at Stanford University have been working hard in the medical field.

In November 2017, the team came up with a new technology called CheXNet. The researchers said that the new technology has surpassed human professional physicians in the accuracy of identifying diseases such as pneumonia in chest X-rays.

The researchers said that the CheXNet algorithm they developed is a 121-layer convolutional neural network that can detect pneumonia through chest X-rays at a level that exceeds that of professional radiologists.

The network is trained on the largest open chest X-ray dataset, ChestX-ray14, which contains 100,000 front view X-ray images of 14 diseases.

As the industry is rushing to lay the groundwork for "AI lung disease diagnosis", some people have questioned the effectiveness of AI imaging in diagnosing lung diseases.

Leifeng.com learned that in a study titled "Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis" published in Nature magazine on May 23, 2016, researchers stated that the sensitivity of the deep learning algorithm reached 100%, but the false positive rate was as high as 40%.

The false positive rate (FPR), also known as the misdiagnosis rate or type I error, is the percentage of people who are actually disease-free but are diagnosed as having a disease based on screening.

This means that AI, as an auxiliary screening system for lung nodules, may still require secondary review by doctors.

Some industry insiders said that due to the high false positive rate, doctors spend too long on the second review. Not only do they need to determine whether the AI ​​has missed the diagnosis, but they also need to rule out the existence of false positives. This does not save them time but increases their tasks.

Samsung's foldable phone is released, priced at 13,000 RMB; Meituan responds to "Dianping will disappear"

Amazing! Samsung releases Galaxy Fold foldable phone, S10 series becomes supporting role

Is Huawei really not going to make cars?

Xiaomi Mi 9 is here, Lei Jun: This is the last time to set the price below 3,000

Looking ahead to the next 16 years in the Guangdong-Hong Kong-Macao Greater Bay Area, how will the industry make its move?

Apple's foldable iPhone patent exposed; JD Finance responds to App issues: wrong development