AI-Based Eye Image Screening Ensures 100% Accurate Diagnosis of Childhood Autism

Scientists captured images of children’s retinas and employed a deep learning AI algorithm for autism diagnosis, achieving a perfect 100% accuracy. These results endorse the use of AI as an unbiased screening tool for early detection, particularly in cases where access to a specialized child psychiatrist is restricted.

The optic disc, where the retina and optic nerve converge at the back of the eye, serves as a gateway to the brain. Recognizing its accessibility and non-invasiveness, researchers are leveraging this structure as a valuable source of crucial information about the brain.

Recently, UK scientists devised a non-intrusive method for swiftly detecting concussions by directing an eye-safe laser onto the retina. Now, researchers at Yonsei University College of Medicine in South Korea have introduced a technique for diagnosing autism spectrum disorder (ASD) and gauging symptom severity in children using retinal images analyzed by an AI algorithm.

Training the Convolutional Neural Network for ASD Screening and Severity Assessment

The study involved 958 participants, averaging 7.8 years in age, whose retinas were photographed, resulting in a total of 1,890 images. Half of the participants had a confirmed ASD diagnosis, while the other half served as age- and sex-matched controls. Assessment of ASD symptom severity utilized scores from the Autism Diagnostic Observation Schedule – Second Edition (ADOS-2) and the Social Responsiveness Scale – Second Edition (SRS-2).

A deep learning algorithm, specifically a convolutional neural network, underwent training using 85% of the retinal images along with symptom severity test scores to develop models for the screening of autism spectrum disorder (ASD) and assessing ASD symptom severity. The remaining 15% of images were reserved for testing.

During ASD screening on the test image set, the AI demonstrated the ability to identify children with an ASD diagnosis, achieving a mean area under the receiver operating characteristic (AUROC) curve of 1.00. The AUROC scale ranges from 0 to 1, where 0.0 signifies predictions that are 100% wrong, and 1.0 signifies predictions that are 100% correct. Notably, even when 95% of the less critical image areas (excluding the optic disc) were removed, there was no significant decrease in the mean AUROC.

Training the Convolutional Neural Network for ASD Screening and Severity Assessment

The researchers stated, “Our models exhibited promising performance in distinguishing between ASD and TD [typical development] using retinal photographs, suggesting that retinal alterations in ASD could serve as potential biomarkers.” They also highlighted the crucial role of the optic disc area, as the models maintained a mean AUROC of 1.00 using only 10% of the image containing the optic disc, emphasizing its importance in distinguishing ASD from typical development.

The mean AUROC value for assessing symptom severity was 0.74, falling within the ‘acceptable’ range, as AUROC values between 0.7 and 0.8 are considered acceptable, while 0.8 to 0.9 is deemed ‘excellent.’

The researchers noted, “Our findings indicate that retinal photographs could offer additional insights into symptom severity.” They observed that effective classification was achievable only for ADOS-2 scores and not for SRS-2 scores. This discrepancy may arise from the fact that ADOS-2 assessments are conducted by trained professionals with sufficient time for evaluation, whereas SRS-2 assessments are typically completed by caregivers in a shorter timeframe. Consequently, the former is likely to more accurately reflect one’s severity status than the latter.

The study included participants as young as four years old. The researchers suggest that their AI-based model could serve as an objective screening tool from this age onward. However, since the newborn retina continues to develop until the age of four, further research is necessary to determine the tool’s accuracy for participants younger than that age.

“While future studies are needed to establish generalizability, our study marks a significant step toward developing objective screening tools for ASD, addressing pressing issues such as the limited accessibility of specialized child psychiatry assessments due to resource constraints,” concluded the researchers.

Read the original article on: New atlas

Read more: New Research Discovers Links Between Genetic Factors in Autism Spectrum Disorder (ASD)