Deep Learning Could Lead to Earlier Turner Syndrome Diagnosis

Blog Banner

As part of its mission to support the Turner Syndrome (TS) community, the Turner Syndrome Foundation (TSF) is committed to making earlier diagnosis a reality. Early diagnosis means that TS patients and their caregivers can better prepare for any related medical issues and take advantage of available treatments. This article discusses a Chinese study that used deep learning, a type of powerful computer programming that classifies objects with “neurons,” to detect facial characteristics of patients with TS.

The Importance of Early Diagnosis

Early diagnosis is crucial because it helps:

  • allow doctors to start estrogen replacement and/or growth hormone therapies early enough improve the health and height of TS patients; and
  • make potentially life-saving medical interventions for heart, liver, and other health challenges possible, preventing them from developing into dangerous problems as time passes.

Despite the fact that scientists have known about TS for decades, patients still often receive late diagnoses or misdiagnoses. That is why it is crucial for both doctors and patients to spread TS awareness. 

Note: Please see the Glossary below for an explanation of some of the terms in this article.

"Investing in science education and curiosity-driven research is the future."

Purpose of the Study

This research study attempted to design accurate facial recognition software to detect TS cases based on common facial features to create a refined deep learning program that might help diagnose TS earlier.

The Researchers

Researchers from the Chinese Academy of Science (CAMS) in Beijing, China included: 

  • Zhen Shen;
  • Yin Bao, who also works with the University of Southampton’s School of Electronics and Computer Science in the UK;
  • Lulu Niu, who also works with CAMS’ School of Artificial Intelligence;
  • Xisong Dong;
  • Xiuqin Shang; and
  • Gang Xiong, who also works with CAMS’ Guandon 3D Printing and Intelligent Manufacturing Engineering Research Center and Cloud Computing Center in Dongguan, China.

Researchers from Peking Union Medical College’s Hospital (PUMCH), CAMS, and PUMC’s Endocrine Key Laboratory in Beijing included: 

  • Huijuan Zhu,
  • Siyu Liang,
  • Shirui Wang,
  • Xiangying Li,
  • Shi Chen, and
  • Hui Pan.

Zhouxian Pan, from PUMCH and the CAMS Allergy Department in Beijing, also participated in the study.

Scope of the Study

Peking Union Medical College Hospital (official website)

The researchers collected photographs between July 12, 2016 and April 16, 2019 from patients at PUMCH’s Short Stature Clinic. They then collected photos from first-time patients from April to May 23, 2019.

The Beijing Municipal Natural Science Foundation, the National Natural Science Foundation of China, the CAMS Initiative for Innovative Medicine, the National Natural Science Foundation of China, and the CAS Key Technology Talent Program funded the study. The researchers submitted the article for publishing in 2019.

How the Study Was Conducted

After PUMCH approved the project, the group obtained consent for the patients to participate in the study. They then set out to determine whether a deep learning program could identify any missed cases of TS. The researchers then collected and analyzed the data via the following method:

  • First, they collected photographs of the TS patients and first-time patients with neutral facial expressions. Some patients had multiple photos taken until they reached puberty.
  • They also gathered patient age, height, weight, treatment, and karyotype information during this time. 
  • They then excluded study samples, including first-time patients diagnosed with TS after karyotyping, patients who had no face-altering genetic conditions, and patients with poor photos.
  • Information from the remaining 170 TS patients and 1,070 first-time patients was then imported into the software that used deep conventional learning networks (deep CNN)
  • The deep CNN classified and scored images of patients, with higher scores meaning an image was similar to thata of a TS patient. 
  • Samples were then used to train and test the program in seven different scenarios. Each scenario involved changing the first or last photograph of TS patients or control subjects and classifying by height or age matching. 
  • The researchers conducted second test with a smaller sample size of two TS patients and 35 control subjects. This test determined whether the researchers’ software would clinically identify TS patients.
  • Finally, the group performed statistical analysis to evaluate the system’s accuracy, the specificity and sensitivity of the second test, and how age and height affected the results.

Research Results

Using this algorithm, the researchers made interesting discoveries, such as:

  • For all seven scenarios, the classification system’s predictions of diagnosis were between 93 and 96% correct.
  • The second test’s results demonstrated that the software’s average sensitivity (96.7%), specificity (97%), and accuracy (96.9%) were high, displaying its capabilities in identifying TS patients in a clinical setting.
  • The classification system was also highly accurate, demonstrating specificity and sensitivity levels greater than some of the authors’ previous attempts.
  • It was found that in the age range used (3 to 30 years old), age matching did not significantly affect the results.
  • Height matching with the range used (80 to 185 cm, or 31.5 to 72.8 in) did not affect the results.

Other Studies

The researchers mentioned two sources that seemed to confirm their findings that facial recognition programs, and TS-specific ones, could accurately identify genetic disorders like TS. 

A study led by Wenai Song, ”Multiple facial image features-based recognition for the automatic diagnosis of turner syndrome,” successfully used automatic facial recognition software to detect TS. Its average accuracy was between 83.4 and 84.6%.

Another study published in Nature Medicine conducted by Yaron Gurovich, et al titled “Identifying facial phenotypes of genetic disorders using deep learning” demonstrated 91% accuracy in identifying over 200 genetic disorders using a deep learning facial recognition software. Amazingly, it even beat this study in recognizing the disorders.

However, both studies suggested that programs must continue to be refined for future use and clinical application. They also suggested that study sample sizes must be larger to ensure accuracy.

Study Limitations

Limitations of this study included: 

  • Not all of the database’s and second test’s controls were genetically tested. Therefore, some control subjects could have TS, affecting the results.
  • The study only consisted of one database and did not create age or height categories, which could affect the results because photographs of children and adults were compared.

The above factors suggest that the algorithm needs more refining to encompass all potential TS symptoms, while still remaining specific enough to make it practical.

The researchers stated there were no conflicts of interest.

Suggestions for Future Research

The authors suggested that a future study involving a larger population is needed further to determine the program’s accuracy in a clinical setting. They also suggested that clinical use of this technology would require separate databases for children and adult patients to improve accuracy.

Importance for the TS Community

This information is valuable for the TS community because it can be difficult for individuals with TS to receive an early diagnosis, which is crucial for an early start of live-saving medical treatments that can improve quality of life. The article’s findings present a potential solution to late diagnosis, helping to ensure that TS patients receive quick, easy, and early diagnoses through facial recognition software.

You can help promote TS awareness by:

  • joining TSF’s educator membership and checking out TSF’s resources to learn more about TS and the TS community and how to support students with TS, 
  • participating in TS research or collaborating with TSF for your research project to help the TS community, and/or
  • joining TSF’s professional membership to learn how to support the TS community in obtaining the medical services they need.

Glossary

  • Deep Learning is a type of powerful computer programming that classifies objects without the help of outside computer software by using multiple “neuron” layers.
  • DCNN (Deep Conventional Neural Networks) are powerful computer programs that function like the human brain. They classify data by using “neurons” that take thousands of small data pieces and categorize them to a  certain neuron channel’s number. After that, they continue going through the channels. Then, the programs determine whether the data’s sums of the channel numbers complete the second to last neuron’s equation–the activations equation. If it is completed, then the equation passes the data into the program’s categories. As the programs analyze more data, the classifications change.
  • Karyotyping is when a photo of a patient’s chromosomes in the blood is studied to see if a patient has a genetic condition.
This photo from DLT Labs displays how deep learning programming differs from machine learning.

Takeaways & Action Steps

  • Early detection of TS is important because allows an earlier start to treatments that improve health outcomes and potentially save lives.
  • This study’s researchers used deep learning CNN to create a successful, sensitive, and accurate classification system that could distinguish TS patients from other patients.
  • However, before this program can be used widely, more research with larger samples and more refined categories is needed.
  • This technology would provide a simple and effective way of diagnosing TS, which would help prevent delayed diagnoses or misdiagnoses.
  • If you want to help increase early TS diagnosis, please support TSF by spreading awareness about TS, advocating for the TS community, and/or holding a fundraiser that supports TSF. 

Written by Julianne Franca, TSF volunteer blog writer. Edited by Ruchika Srivastava, TSF volunteer blog editor, and Susan Herman, TSF volunteer lead blog editor.

Sources

Clinical

Non-Clinical

TSF Resources

 © Turner Syndrome Foundation, 2021

Leave a Reply

©2023 Copyright Turner Syndrome Foundation - All rights Reserved.

X
%d