Cornelia de Lange syndrome is very rare, affecting only about one in about 45,000 people. The disease is often associated with mental retardation. The syndrome is also noticeable on the outside: those affected usually have dense, curved eyebrows that have grown together over the root of the nose, a large, flat nose with forward nostrils, a narrow upper lip, downturned corners of the mouth and a rather small chin.
So far, doctors have mainly diagnosed the disease based on the external appearance of the patients. Because the disease is so rare, it's not always easy. Now researchers have developed a software that can reliably detect rare hereditary diseases based on portrait photos, especially of children.
Software detects 200 hereditary diseases
The program called DeepGestalt developed scientists from the USA, Israel and Germany, as they report in the journal "Nature Medicine". It should be able to recognize more than 200 usually extremely rare syndromes, which often cause problems even in childhood. If there is an initial suspicion, the software limits the number of possible genetic causes and thus accelerates the diagnosis, says Peter Krawitz from the University of Bonn.
If you take all rare hereditary diseases together, they often occur: About two to eight percent of the population have a genetic syndrome, says Krawitz. One-third to one-half of these illnesses are associated with mental impairments, which often show up even in infancy.
"Because of the large number of possible syndromes and their rarity, the correct diagnosis is a lengthy and expensive process," writes the team around Yaron Gurovich of the Boston-based company FDNA. So far only a few experts could recognize unusual appearances or extremely rare symptoms.
For those affected and their families, this often means a doctor's odyssey for years to decades, the constant search for the cause of the problems, new therapeutic attempts, shattered hopes. Intelligent programs could significantly improve the situation, according to the researchers.
The DeepGestalt software examines faces for characteristic abnormalities and then analyzes the shape of the eyes, mouth, chin or the distance between the eyebrows. Overall, the face recognizes 130 points in the face and compares them to 216 syndromes.
It works like a neural network that recognizes certain patterns in the face. "In the database, the patient photo is compared with many images and a total similarity is determined," says bio-computer scientist Krawitz.
Training with more than 17,000 pictures
According to Krawitz, the training of the software on a data set of more than 17,000 images was the most time-consuming. Thus, the program practiced the detection of Cornelia de Lange syndrome based on 614 images of those affected and nearly 1100 images of other people. In a subsequent test, whether or not someone has this syndrome, the program achieved a reliability of 97 percent.
In the case of Angelman syndrome, which used images from just under 770 people and nearly 2,700 other people, reliability was 92 percent. Those affected often have a small head that is flattened at the back, and a rather large mouth with protruding upper jaw.
Initially, these two tests were all about whether someone has this syndrome or not. In two further tests, the researchers then examined how well DeepGestalt can associate a facial photo with one of 216 different genetic defects.
Program provides only suspected diagnosis
After analysis, the system issued a top ten list of possible diagnoses: the likelihood that the actual genetic defect was among those top ten was about 90 percent. In about 65 percent of cases, even the most likely diagnosis occurred.
The software could be used by pediatricians, for example, to whom parents with conspicuous children would come, says Krawitz. However, the program only supplies suspected diagnoses, which then have to be checked by laboratories. In addition, only those genetic defects can be detected, which also cause significant morphological deviations of the face shape - not all hereditary diseases bring with it.
"The value lies in the fact that paediatrists can ideally initiate a targeted diagnosis in consultation with a human geneticist," says the bioinformatician. Thus, even experts are often overstrained with the diagnosis, because the number of newly discovered syndromes continuously increase.
Recognizing hereditary diseases is just one of many applications of artificial intelligence in medicine. According to Krawitz, similar software is being developed for the evaluation of other images, such as MRI images or photos of the retina.