Diagnostic performance of deep learning models using parotid gland ultrasonography in Sjögren’s Disease

Ünal M., Türk E., Danışman T., Öğüt T. S., Bosnalı B., Erbasan F., ...Daha Fazla

BIOMEDICAL SIGNAL PROCESSING AND CONTROL, cilt.116, sa.1, ss.109615, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 116 Sayı: 1
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1016/j.bspc.2026.109615
  • Dergi Adı: BIOMEDICAL SIGNAL PROCESSING AND CONTROL
  • Derginin Tarandığı İndeksler: Scopus, Science Citation Index Expanded (SCI-EXPANDED), Compendex, EMBASE
  • Sayfa Sayıları: ss.109615
  • Akdeniz Üniversitesi Adresli: Evet

Özet

Objectives: Sjögren’s Disease (SjD) is a chronic autoimmune disorder that affects the salivary glands, often requiring invasive diagnostic procedures such as labial salivary gland biopsy. Although ultrasonography provides a non-invasive alternative, its interpretation remains highly operator dependent. This study aimed to evaluate lightweight deep learning models for the automated diagnosis of SjD using parotid gland ultrasonography images.

Material/Methods: A total of 123 patients with SjD and 123 age- and sex-matched healthy controls were included. Bilateral grayscale parotid ultrasonography images were analyzed using a deep learning framework. Multiple convolutional neural network architectures, including YOLO11 variants, ResNet, DenseNet, and EfficientNet, were systematically compared. Model training and evaluation were performed using a repeated five-fold grouped cross-validation strategy to maintain subject-level independence and prevent data leakage.

Results: Among the tested architectures, the YOLO11s-cls model achieved the highest overall performance, with a mean accuracy of 93.1 %, precision 93.8 %, sensitivity 92.4 %, specificity 93.8 %, and F1 score 93.0 %. Gradient-based activation maps confirmed that the network focused on diagnostically relevant glandular regions.

Conclusion: The lightweight deep learning models enable accurate, objective, and reproducible diagnosis of SjD using only parotid ultrasonography images. These findings highlight the potential of artificial intelligence in ultrasonography to standardize image interpretation and reduce operator dependence in clinical practice.