New Research Links Retinal Imaging and Deep Learning to Improved Schizophrenia Detection

A groundbreaking study led by Pawel Karczmarek, PhD, has revealed that deep neural networks and advanced aggregation functions, particularly new generalizations of the Choquet integral, may enhance the classification and detection of schizophrenia through retinal imaging.

Retinal Biomarkers in Schizophrenia

Previous research has established that specific retinal variables can serve as biomarkers for neurodegenerative diseases. Notably, studies have confirmed that individuals with schizophrenia tend to have thinner retinal layers, supporting the potential role of retinal imaging in disease detection.

Study Objectives and Methods

The primary goal of the study was to evaluate how aggregation-based classification methods perform in distinguishing individuals with schizophrenia from healthy controls. The research team tested over 300,000 different aggregation operator variants, identifying only a few that significantly improved classification accuracy.

Additionally, this study marks the first application of classical and deep learning classifiers—combined with different aggregation operator variants—to OCT (optical coherence tomography) data in schizophrenia research.

Study Design and Participants

       • The study included patients diagnosed with schizophrenia based on ICD-10 classification (code F20.x), hospitalized at the Medical University Psychiatric Clinic in Poland.

       • The chlorpromazine equivalent dosage for participants was 452.22 mg/day (SD = 128.2).

       • 57% of patients received only atypical antipsychotic drugs, while 43% received a combination of classic and atypical medications.

       • A healthy control group was also included for comparison.

Retinal Imaging and Data Processing

       • Retinal imaging was performed using the Optopol Copernicus REVO SD-OCT, with measurements generated via Optopol SOCT version 11.0.7.

       • Strict exclusion criteria included ophthalmic diseases, diabetes, untreated hypertension, history of eye trauma or surgery, neurological head injuries, extreme refractive errors (>±5 diopters), glaucoma risk (DDLS scale ≥6), history of psychoactive substance addiction (except nicotine), inherited intellectual disabilities, and dementia.

       • Each patient underwent two OCT scans per eye, totaling four images per participant.

The final dataset included 59 schizophrenia patients and 61 healthy controls. Data processing was conducted using Python-based custom programs, incorporating Keras, Scikit-Learn, and TensorFlow libraries.

Key Findings: Choquet Integral and Deep Learning

The study evaluated various OCT features, including:

       • Macular thickness

       • Retinal macular volume

       • Outer nuclear and outer plexiform layers

       • Macular and peripapillary retinal nerve fiber layers thickness

       • Ganglion cell complex

These features were analyzed using deep neural networks combined with an aggregation procedure, with one of the most effective being the Choquet integral. This method not only weighs individual classifiers but also assesses how well their results align, making it more than just a simple averaging tool.

A Smarter Approach to Disease Classification

The study authors explained:

“The Choquet integral functions like a team of experts working together, each specializing in a different area. The final decision is more precise because it considers both individual opinions and the level of agreement among classifiers, ensuring a more trustworthy outcome.”

Implications and Future Research

The integration of retinal imaging, deep learning, and advanced aggregation functions represents a promising direction for schizophrenia detection and classification. While further validation and clinical trials are needed, this research highlights the potential of AI-driven ophthalmic diagnostics in identifying neurological disorders through non-invasive retinal analysis.

Reference:

Karczmarek P, Plechawska-Wójcik M, Kiersztyn A, et al. On the improvement of schizophrenia detection with optical coherence tomography data using deep neural networks and aggregation functions. Sci Rep. 2024;14:31903. https://doi.org/10.1038/s41598-024-83375-7