Evaluating a General-Purpose AI Model for Diagnosing Vocal Fold Lesions Using Static Laryngeal Images

Gojayev E. K., Büyükatalay Z. Ç.

Turkish Journal of Ear Nose and Throat, cilt.35, sa.4, ss.171-177, 2025 (Scopus, TRDizin)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 35 Sayı: 4
  • Basım Tarihi: 2025
  • Doi Numarası: 10.26650/tr-ent.2025.1750893
  • Dergi Adı: Turkish Journal of Ear Nose and Throat
  • Derginin Tarandığı İndeksler: Scopus, TR DİZİN (ULAKBİM)
  • Sayfa Sayıları: ss.171-177
  • Anahtar Kelimeler: Artificial intelligence, diagnostic imaging, laryngeal diseases, laryngoscopy, vocal cords/pathology
  • Ankara Üniversitesi Adresli: Evet

Özet

Objective: To evaluate the diagnostic performance of a general-purpose artificial intelligence (AI) model in classifying vocal fold lesions using static laryngeal images. Materials and Methods: This retrospective study included 175 cases representing 14 vocal fold pathologies. Two static endoscopic frames per case—captured during inspiration and phonation—were analysed using a GPT-4-based AI model via structured diagnostic prompts. The model had no prior training on laryngeal images. The diagnostic accuracy, sensitivity, specificity, precision, and F1-score were calculated. Chi-square testing was used to compare the observed accuracy to chance. Results: The overall diagnostic accuracy was 29.14%. The model showed perfect accuracy (100%) in vocal fold haemorrhage and chronic fungal laryngitis, but failed to identify vocal fold paralysis and leukoplakia. The sensitivity ranged from 0% to 100%, while the specificity was more stable (66%–75%). The macro-average and weighted-average F1-scores were 33.38% and 29.14%, respectively. The model performed significantly better than chance (p<0.001), with substantial variation across diagnoses. Conclusion: Although the performance was inconsistent across pathologies, the model demonstrated high diagnostic accuracy in selected lesions. These findings support the potential of AI-assisted tools in laryngeal diagnostics, while underscoring the need for domain-specific training and validation.