Release date: 2016-01-08 In recent years, the performance of "image recognition technology" that recognizes an object from an image has been rapidly improved by "deep learning." Enlitic, a startup based in San Francisco, USA, applies deep learning to the detection of malignant tumors such as cancer. The system developed by the company has a higher cancer detection rate than radiologists. Deep learning is a machine learning method that uses a "deep neural network" that simulates the human brain structure. It can also be used for speech recognition and natural speech processing, but it has achieved significant results in the field of image recognition. Source: Nikkei BP Nitrile Coated Glove,Nitrile Coated Work Gloves,Nitrile Coated Hand Gloves,Nitrile Foam Coated Gloves Jiangsu Hespax Security Co., Ltd , https://www.hespax.com
In the competition for testing the performance of image recognition technology, the "ILSVRC" (ImageNet Large-Scale Visual Recognition Challenge) using the image database "ImageNet" is the most famous. In the 2015 ILSVRC, which was attended by well-known IT companies such as Google, Intel, Qualcomm and Tencent, Microsoft Research won. The test content of this competition is whether it can accurately classify 1000 pictures. Microsoft's classification error rate is only 3.6%.
In terms of the minimum classification error rate of the past competitions, it was 7.4% in 2014, 11.1% in 2013, and 15.3% in 2012. In this event, the team with deep learning won in 2012. At that time, the classification error rate of 15.3% was already "shocking". Just three years later, the winning Microsoft Research Institute proposed that the recognition accuracy of image recognition technology based on deep learning can exceed human precision.
However, Enlitic's data analyst Rewon Child (Figure 1) said, "ImageNet's competition only recognizes whether the object on the 224-pixel × 224-pixel image is a cat or a dog. This task is not difficult." He also pointed out that "we are challenging more difficult image recognition." 
The more difficult image recognition Enlitic needs to challenge is to find malignant tumors such as cancer from images such as X-ray, CT scan, ultrasound, and MRI. Child explained: "The resolution of X-ray photos is 3000 pixels vertically × 2000 pixels horizontally. The size of the malignant tumor is 3 pixels vertically × 3 pixels horizontal. A small shaded object is judged from a very large image. Whether it is a malignant tumor is a very difficult task."
Image recognition software for finding malignant tumors from X-ray photographs and CT scan images was developed using one of the deep learning methods "Convolutional Neural Network (ConvNet, Convolutional Neural Network)". ConvNet conducts machine learning on a large number of medical image data that have been examined by radiologists for the presence of malignant tumors and tumor locations, and automatically summarizes "characteristics" that represent the shape of malignant tumors and "models" that emphasize the characteristics of whether or not to detect malignant tumors. ConvNet will apply the patterns found to the new medical image to see if there are malignant tumors in the image.
According to Enlitic, the company developed a malignant tumor detection system that is more accurate than a radiographer. Enlitic used the lung cancer related image database "LIDC (Lung Image Database Consortium)" and "NLST (National Lung Screening Trial)" to verify that the lung cancer detection accuracy of the system developed by the company was higher than that of a radiologist. The accuracy is more than 50%.
Enlitic will provide radiologists with a malignant tumor detection system (Figure 2). In the United States, radiographers are employed by medical imaging diagnostic services companies and medical institutions that will become Enlitic's customers. In October 2015, Australian medical imaging diagnostics company Capitol Health announced the adoption of Enlitic's system. This is the first time Enlitic's system has been adopted. At the same time, Capitol Health invested $10 million in Enlitic. 
Child said, "The radiographer diagnoses a patient's CT scan image takes 10 to 20 minutes, and the diagnosis report takes about 10 minutes. If the company's system is used, the CT scan image diagnosis time can be halved." He also predicted, "Although image recognition technology can be used to determine whether there is a malignant tumor, but in view of government regulations, medical institutions are not likely to use radiographers. However, if the work efficiency of radiographers is doubled, it is developing. Patients in the country can use CT scans more conveniently."
Excellent application developed by medical laymen
About Enlitic, the author is very interested in the company's members. According to Child, the company's data analysts are all people without medical work experience. Child himself is a political science major at Yale University in the United States. He has studied the use of statistics and other methods to analyze the "measurement social science" of society.
The company's data analysts are all recruited from the "Kaggle" website. Kaggle is a "data forecasting competition" website that provides data analysis related topics to data analysts around the world and compares the research results. The project has a bonus, and the data analyst can solve the problem and get a bonus.
The subject of Kaggle is to entrust data analysis to external companies and companies looking for excellent data analysts. Enlitic raised the topic at Kaggle and hired an analyst who solved the problem perfectly. It has nothing to do with medical knowledge.
Why Enlitic chose Kaggle In fact, Enlitic founder and CEO Jeremy Howard was also the president and chief analyst of Kaggle. In other words, Enlitic is also a company that knows Howard, a good data analyst who can find good data analysts through Kaggle, to take advantage of data analysts to start new businesses.
Data analysts have begun to change the industry
Child said, "Howard wanted to create a company that would keep a good data analyst working for 25 years, so he founded Enlitic. I heard that his business, in addition to medical image diagnosis, also considered looking for oil and natural gas." Data analysts are at the forefront before there are application areas.
Child said, "Kaggle gives everyone the opportunity to be recognized as long as they can write great code. This is a very fair and democratic place." Data analysts gathered here have developed applications in areas where there is no “business knowledge†and have begun to achieve results beyond existing insiders. Through Enlitic, the company is not only able to predict the medical industry, but also predict the future of the entire industry. In this sense, the company is also very important.
Figure 1: Rewon Child, Enlitic Data Analyst
Figure 2: Schematic diagram of the detection of malignant tumors using Enlitic's system Source: Enlitic, USA