Don't Panic When Acne Appears on Your Face, Let Artificial Intelligence Help

2019-10-22

Artificial intelligence has already begun to play a role in many areas of the medical industry, such as ophthalmology and internal medicine, and has achieved great success. Since 2016, there have been many studies on the application of AI in dermatology. In fact, skin problems have become one of the most common diseases in the world, following colds, fatigue, and headaches. Globally, about 1.9 billion people suffer from skin diseases at some point in time. In the United States alone, 37% of patients visiting clinics have at least one skin problem. In China, there are also over 200 million skin disease patients. However, more than half of these patients are treated by non-dermatologists. (Think about whether you have ever squeezed a pimple yourself...) Due to the shortage of dermatologists worldwide, many patients cannot receive professional advice and treatment, and general practitioners are not as accurate in determining the condition. Therefore, Google's Deep Learning System (DLS), an AI dermatologist, was developed. This artificial intelligence system from Google can provide diagnoses with the same accuracy as dermatologists for 26 types of skin diseases. Previous machine learning studies on skin diseases have typically focused on early detection of skin cancer, particularly the identification of malignant and benign skin cancers. However, Google also pointed out that over 90% of skin diseases are not malignant, and solving these common diseases would greatly help alleviate global skin diseases. Google software engineer Yuan Liu and Google Health technical program manager Dr. Peggy Bui stated: "We developed the Deep Learning System (DLS) to address the most common skin problems in primary care. This research highlights the potential of DLS to enhance the diagnostic capabilities of general practitioners without additional specialized training." Google noted that clinical cases are not always determined with absolute certainty from the beginning. Doctors provide a differential diagnosis, a list of possible conditions, which can be used for additional examinations until the condition is confirmed. DLS mimics this idea by providing a ranked list of possible skin diseases for patients, helping with patient classification and subsequent diagnosis and treatment. The training data for DLS includes one or more clinical images of skin abnormalities, as well as metadata such as age, gender, and symptoms. Google used a total of 17,777 de-identified cases as training data, with cases from 2010 to 2017 used for training and cases from 2017 to 2018 used for evaluation. During the training of DLS, over 40 dermatologists provided more than 50,000 differential diagnoses. To evaluate the accuracy of DLS, Google compared its results with diagnoses from three dermatologists. In 3,756 cases, the probability of DLS correctly diagnosing the skin condition in the top-ranked list was 71%, and the accuracy of the top three diagnoses was as high as 93%. The diagnostic accuracy of DLS is comparable to that of dermatologists, averaging 75%, and far exceeds that of primary care physicians (60%) and nurse practitioners (55%). The accuracy of the top three diagnoses can even help dermatologists determine the condition. Since skin diseases are highly correlated with skin types, visual assessment of the skin itself is crucial for diagnosis. Therefore, to evaluate potential biases towards skin types, the team tested the performance of the AI system based on the Fitzpatrick skin typing, which ranges from type I (pale, always burns, never tans) to type VI (darkest brown, never burns). They focused on skin types Fitzpatrick II to IV, which account for at least 5% of the data, and found that the accuracy of DLS was similar in these categories, with accuracy ranging from 69% to 72% for the top-ranked diagnosis and 91% to 94% for the top three diagnoses. The researchers attributed the overall accuracy of this system to the presence of metadata in the training corpus and stated that the results of the study suggest that their approach may "prompt clinicians to consider potential factors" that were not initially used for differential diagnosis. However, they also pointed out that their training corpus was only derived from a remote teledermatology service. Some Fitzpatrick skin types were too rare in their dataset to be meaningfully trained or analyzed, and certain skin conditions, such as melanoma, could not be accurately detected due to a lack of available data samples. Liu and Bui wrote: "We believe that adding more histopathologically confirmed cases of skin cancer to the training and validation sets can address these limitations." In foreign countries, there have been previous examples of AI-assisted diagnosis of skin diseases. For example, the Israeli company DermaCompare developed an application that detects melanoma through images. Users take "total body photography" using their phone's camera and upload it to the app, which can then diagnose whether a mole is a sign of melanoma. In 2017, a team led by researchers from Stanford University in the United States used 130,000 images related to skin lesions to "train" an AI program developed by Google to classify and identify whether skin lesions are benign or malignant. The model's performance in skin cancer identification was comparable to that of 21 dermatologists. In 2018, a research team from Germany, the United States, and France trained an AI system with over 100,000 images to distinguish between dangerous skin lesions and benign skin injuries. This machine, a deep learning convolutional neural network (CNN), later outperformed 58 dermatologists from 17 countries in a test to differentiate between photos of malignant melanoma and benign moles. More than half of the dermatologists were "expert" level with over 5 years of experience, 19% had 2 to 5 years of experience, and 29% were beginners with less than 2 years of experience. "The CNN performed better than most dermatologists," the research team wrote in a paper published in the Annals of Oncology. On average, human dermatologists accurately detected 86.6% of skin cancers, while the CNN's accuracy was 95%. The first author of this study, Holger Haenssle from Heidelberg University, said in a statement, "The CNN detected more melanomas, which means it had higher sensitivity than dermatologists." However, it also "misdiagnosed a small number of malignant melanomas, leading to unnecessary surgeries." The performance of these "dermatologists" improved as they received more information about patients and their skin injuries. The research team stated that artificial intelligence could be a useful tool for faster and easier diagnosis of skin cancer and for surgical removal before it spreads. They added that there are approximately 232,000 new cases of melanoma worldwide each year, resulting in 55,500 deaths. In China, there have also been significant advancements in AI applications in dermatology. For example, Xiangya Second Hospital of Central South University, in collaboration with Dingxiangyuan and DANA Technology, has developed the first AI-assisted diagnostic system for skin diseases and held a press conference. The system currently focuses on a series of diseases such as lupus erythematosus and dermatitis, with an accuracy rate of over 85%. In addition, other dermatology departments in Chinese hospitals have gradually started using AI diagnostic tools. For example, Peking Union Medical College Hospital and Beihang University have begun using automatic recognition of dermatoscope images, which was demonstrated at a recent skin imaging continuing education class. Wuhan Union Hospital has also collaborated with a company from Hong Kong to apply their developed intelligent skin detection system (Dr.Skin), which can effectively diagnose common skin diseases through image analysis. The China Skin Image Database (CSID) project initiated by Professor Yong Cui from the China-Japan Friendship Hospital aims to establish a learning resource specific to the Chinese population for auxiliary diagnostic models. It is also an important learning resource for AI-based intelligent diagnosis of skin diseases. From a broader perspective, AI diagnostic systems for skin diseases have two main applications: doctors and patients. For doctors, AI diagnostic systems can be excellent assistants, improving efficiency and reducing misdiagnosis rates. For patients, they can quickly gain preliminary understanding of their own conditions through user-friendly operations. Currently, AI as an auxiliary diagnostic tool for doctors is the main scenario for AI applications in dermatology, especially for primary care physicians with weaker diagnostic abilities. However, it seems inevitable for AI-assisted diagnosis to move towards the patient side and enable self-examination.

RECENT AI NEWS

RECENT AI TOOLS