Rare diseases: diagnosis by photo
The neural network was taught to recognize 216 rare hereditary diseases by photo
Ekaterina Rusakova, N+1
Researchers have developed an artificial intelligence system that allows to diagnose 216 rare hereditary diseases using photographs with high accuracy. As reported in Nature Medicine (Gurovich et al., Identifying facial phenotypes of genetic disorders using deep learning), she was taught to recognize a genetic disorder (choose from the 10 most likely options) with 91 percent accuracy. Scientists have also simplified the application of the system in practice: they have created a mobile application for doctors that allows you to identify a genetic disorder from a patient's photo.
It is often difficult to diagnose a hereditary disease. There are several thousand diseases associated with genetic disorders, most of which are extremely rare. Many doctors may simply not encounter such diseases during their practice, so a reference computer system that would help recognize rare hereditary diseases would facilitate diagnosis. Researchers have already created similar systems based on facial recognition, but they were able to identify no more than 15 genetic disorders so far, while the accuracy of recognition of several diseases did not exceed 76 percent. In addition, such systems sometimes could not distinguish a sick person from a healthy one. At the same time, the training sample often did not exceed 200 photos, which is too small for deep learning.
Therefore, American, German and Israeli scientists and employees of the FDNA company, led by Yaron Gurovich from Tel Aviv University, developed the DeepGestalt facial recognition system, which made it possible to diagnose several hundred diseases. Using convolutional neural networks, the system divides the face into separate fragments with dimensions of 100×100 pixels and predicts the probability of each disease for a particular fragment. Then all the information is summed up, and the system determines the probable disorder for the person as a whole.
DeepGestalt breaks the face in the photo into separate fragments and evaluates how well they correspond to each of the diseases in the model. According to the totality of fragments, the system makes a ranked list of possible diseases (from an article in Nature Medicine).
The researchers trained the system to distinguish a specific hereditary disease from a number of others. For training, they used 614 photographs of people suffering from Cornelia de Lange syndrome, a rare hereditary disease that manifests itself, among other things, in the form of mental retardation and congenital defects of internal organs. As a negative control, the authors used more than a thousand other images. DeepGestalt distinguished Cornelia de Lange syndrome from other diseases with 97 percent accuracy (p=0.01). The authors of other studies managed to achieve 87 percent accuracy, while experts made the correct diagnosis on average in 75 percent of cases. In another experiment, scientists used 766 photographs of patients with Angelman syndrome ("Parsley syndrome"), which, among other things, is characterized by chaotic movements, frequent laughter or smiles. The system recognized the disease with 92 percent accuracy (p=0.05); in the previous study, the detection accuracy was 71 percent.
The researchers also taught the system to recognize different types of the same hereditary disease using the example of Noonan syndrome. There are several types of this disorder, each of which is caused by mutations in a certain gene and each has small differences in facial features (for example, rare eyebrows). Using a sample of 81 photos, the authors of the article taught the DeepGestalt system to distinguish five types of this disease with an accuracy of 64 percent (p<1×10-5).
In total, scientists used a total of 17106 photographs representing 216 hereditary diseases to train the system. The researchers tested the effectiveness of DeepGestalt on 502 photographs of patients who had already been diagnosed, and on another sample of 329 photographs of patients with a known diagnosis from the London Medical Database. The system determined the patient's disease from the 10 most likely variants with an accuracy of 91 percent (p<1×10-6).
The researchers also facilitated the application of DeepGestalt in practice — they created a platform for the diagnosis of hereditary diseases by phenotype, as well as a mobile application for physicians Face2Gene, with which a doctor can diagnose his patient.
Portal "Eternal youth" http://vechnayamolodost.ru