Korean medical researchers have developed a generative AI technology capable of diagnosing Parkinson's disease -- which is often difficult to identify --with an accuracy of up to 99.7 percent and predicting changes in patients’ brain images.
Parkinson's disease, caused by a decline in dopamine-secreting nerve cells in the brain, is the second most common degenerative brain condition after Alzheimer's disease. Common symptoms include tremors, slow movements, and muscle stiffness, though non-motor symptoms such as depression and dementia can also occur.
Early detection is critical. However, in its initial stages, Parkinson's often resembles normal aging or other neurological conditions, making early diagnosis challenging.
An imaging test known as DAT PET (dopamine transporter positron emission tomography) can directly evaluate the status of dopamine neurons. However, it requires highly trained personnel, and the interpretation can be subjective. To address these limitations, a new AI-based technology has been developed.
Asan Medical Center announced Wednesday that a research team led by Professor Kim Nam-kug and Dr. Lee Yoo-jin of the Department of Convergence Medicine, along with Professor Chung Sun-ju of the Department of Neurology, developed an AI model that learns from brain images, generates predictions, and uses this information to diagnose Parkinson's disease.
After being trained on 1,934 DAT PET images, the model demonstrated up to 99.7 percent accuracy in clinical validation, including distinguishing early-stage Parkinson's from essential tremor.
The AI can also predict how a patient’s brain image will evolve over time based on learned data -- allowing physicians to better explain disease progression and guide treatment decisions.
The AI model is based on a “foundation model” -- a general-purpose AI trained on large datasets that can be applied to various diagnostic tasks, disease progression predictions, and prognostic image generation.
This foundation model excels at generating images compared to previous models, using a Hierarchical Wavelet Diffusion Auto Encoder (HWDAE) -- a hierarchical diffusion model-based encoder developed by the research team. This model uses the principle of diffusion to analyze complex brain images in layers and repeatedly adds and restores noise to refine image learning.
The researchers trained the AI on 1,934 DAT PET scans, acquired through the 18F-FP-CIT PET test, and validated its performance in three clinical tasks: distinguishing essential tremor from early Parkinson's, differentiating Parkinson's from multisystem atrophy and progressive supranuclear palsy, and predicting the onset timing of motor symptoms in Parkinson's.
The model achieved classification accuracies of 99.7 percent and 86.1 percent in the first two validations, respectively, and showed an R² correlation of 0.519 (closer to 1 indicates higher accuracy) for predicting motor symptom onset.
The 86.1 percent accuracy rate was particularly notable, as distinguishing Parkinson’s from multisystem atrophy and progressive supranuclear palsy is widely regarded as especially difficult.
To test its generalizability, the AI model was also applied to imaging data from other PET scanners at Asan Medical Center and external hospitals. The model maintained its performance, further validating its clinical applicability.
In addition to diagnosing Parkinson’s through brain imaging, the AI also predicts disease progression and presents the results visually. This can help physicians better communicate with patients and tailor treatment strategies.
"This study is significant because we developed an AI model that not only detects various forms of Parkinson's disease early but also predicts its progression using a diffusion model optimized for image generation. We have also confirmed its clinical applicability," Professor Kim said. "We plan to apply this AI model to other neurodegenerative diseases in the future."
Professor Chung added, "This is a breakthrough that can enhance the accuracy of Parkinson’s disease diagnosis. It also enables the generation of predictive images, which many patients are curious about. We hope to develop this further into a clinical tool that truly benefits patients."