Automated identification of early to mid-stage Parkinson’s disease using deep convolutional neural networks on static facial images

Ni Yang, Jing Liu, Lin Wang, Jiajun Ding, Lingzhi Sun, Xianghua Qi, Ying Lu, Wei YanFrontiers in Medical Technology 2025 · DOI

This study, published in Frontiers in Medical Technology, explores a non-invasive, AI-driven approach to detecting Parkinson’s Disease (PD) in its early-to-mid stages using nothing more than static facial photos. By leveraging deep convolutional neural networks (CNNs), the researchers achieved near-perfect classification accuracy, identifying subtle facial markers—known as hypomimia—that often escape the human eye.

TL;DR

Researchers successfully trained deep learning models to identify early-to-mid-stage Parkinson’s Disease from static facial images with up to 99.67% accuracy. By focusing on "masked face" symptoms (hypomimia), the study demonstrates that lightweight CNNs like ResNet18 and MobileNetV3 can serve as highly effective, low-cost screening tools for remote clinical monitoring.

What Problem Does This Paper Address?

Parkinson’s Disease is the second most common neurodegenerative disorder globally, yet early diagnosis remains a significant hurdle. Current diagnostic methods are often subjective, relying on clinical observations of motor symptoms that usually appear only after more than 50% of dopaminergic neurons have already been lost.

While advanced biomarkers (like cerebrospinal fluid assays or specialized neuroimaging) exist, they are invasive, expensive, and inaccessible for large-scale screening. This paper addresses the need for an objective, affordable, and non-invasive "digital biomarker." Specifically, it targets hypomimia—the reduction of facial expression or "masked face"—which is a common early sign of PD but is difficult for clinicians to quantify consistently.

Methods

The researchers evaluated five classical CNN architectures to determine which was best suited for this specific medical classification task: ResNet18, MobileNetV3, EfficientNetV2, VGG16, and Inception-v4.

The workflow involved:

  1. Preprocessing: Extracting facial regions based on 68 landmarks to remove background noise.
  2. Data Augmentation: Using horizontal/vertical flipping and random rotations ($\pm30^\circ$) to increase the dataset size and improve model robustness.
  3. Fine-Tuning: The models were pre-trained on ImageNet and then fine-tuned on the PD dataset by unfreezing final blocks and applying L2 regularization to prevent overfitting.
  4. Interpretability: The team used Grad-CAM (Gradient-weighted Class Activation Mapping) to generate heatmaps, visualizing exactly which facial regions the models prioritized when making a diagnosis.

Datasets

The study utilized a custom dataset recruited from the Affiliated Hospital of Shandong University of Traditional Chinese Medicine:

  • Participants: 208 PD patients (early-to-mid stage, Hoehn and Yahr scale 0–3) and 1,013 healthy controls.
  • Image Count: 2,000 original standardized facial images (1,000 per group) were selected.
  • Augmentation: The dataset was expanded to 6,000 images to enhance training stability.
  • Split: A standard 8:2 ratio was used for training and testing sets.

Key Insights and Ideas

  • Simplicity Wins: Interestingly, the relatively shallow ResNet18 outperformed the more complex Inception-v4. This suggests that for medical tasks where features are subtle but localized, deeper or more "branchy" architectures may be prone to overfitting or noise sensitivity.
  • Clinical Validation via AI: The Grad-CAM heatmaps consistently highlighted the eyes, lips, and nose as the most discriminative regions. This aligns perfectly with clinical descriptions of PD-related hypomimia, such as reduced blinking and "slackened" lip posture, providing "physiological plausibility" to the model's decisions.
  • Screening vs. Diagnosis: The high recall of MobileNetV3 (99.00%) makes it particularly valuable for screening. In a public health context, a high-sensitivity model is preferred to ensure that potential cases are not missed (minimizing false negatives), even if it requires a secondary clinical follow-up.

Key Results

The models demonstrated exceptional performance across the board, with ResNet18 emerging as the top performer:

  • ResNet18: Achieved an F1-score and Accuracy of 99.67%.
  • MobileNetV3: Followed closely with 98.83% accuracy and a standout 99.00% recall, making it the best candidate for mobile/embedded applications due to its lightweight nature.
  • VGG16: Showed the fastest convergence, reaching optimal performance rapidly.
  • Inception-v4: Was the "underperformer" with 87.42% accuracy, exhibiting signs of instability and potential overfitting on the specific facial feature distribution.

Limitations and Future Work

While the results are promising, the authors acknowledge several constraints:

  • Controlled Environments: The images were taken under uniform lighting with neutral expressions. Real-world performance on "selfies" with cluttered backgrounds or varying lighting remains to be tested.
  • Static vs. Dynamic: Static images miss temporal cues like blinking frequency or micro-expressions. The authors suggest that future work should incorporate 3D CNNs or Transformers to analyze short video segments.
  • Ethnic Diversity: The current cohort was primarily of Asian descent. To be globally scalable, the models must be validated against more diverse datasets to ensure they don't pick up on ethnicity-specific facial features as proxies for disease.
  • Telemedicine Integration: The ultimate goal is to embed these models into smartphone platforms for low-cost, large-scale remote screening in aging populations.

Summary updated

Abstract

CNNs, particularly ResNet18 and MobileNetV3, exhibit significant potential for the automated identification of PD from facial imagery. These models offer promising avenues for developing scalable, non-invasive screening tools suitable for early detection and remote healthcare delivery, providing significant clinical and social value in the context of aging populations.
BibTeX
@inproceedings{yang2025automated,
  author = {Ni Yang and Jing Liu and Lin Wang and Jiajun Ding and Lingzhi Sun and Xianghua Qi and Ying Lu and Wei Yan},
  title = {{Automated identification of early to mid-stage Parkinson’s disease using deep convolutional neural networks on static facial images}},
  booktitle = {Frontiers in Medical Technology},
  year = {2025},
  doi = {10.3389/fmedt.2025.1655199},
}
Figure