Derin Öğrenme Yöntemleri Kullanılarak Ultrason Görüntülerinden Tiroid Nodüllerinin Sınıflandırılması

Öz

Tiroid kanseri, endokrin sistemin en yaygın malignitelerinden biridir ve erken tanı klinik tedavi başarısı açısından kritik öneme sahiptir. Tiroid nodüllerinin değerlendirilmesinde ultrason görüntüleme temel yöntem olmasına rağmen, tanısal yorumlar büyük ölçüde uzman deneyimine bağlıdır. Bu çalışmada, tiroid ultrason görüntülerinde güncel derin öğrenme mimarilerini sistematik olarak karşılaştırmak amacıyla standartlaştırılmış bir kıyaslama çerçevesi sunulmuştur. RegNetY-032, ResNet50, EfficientNet, Inception ve YOLOv8 sınıflandırma varyantlarını içeren toplam 15 model, 5000 görüntüden oluşan TN5000 veri kümesi üzerinde aynı ön işleme ve eğitim koşulları altında değerlendirilmiştir. Performans; Accuracy, Precision, Recall, F1-skoru ve AUROC metrikleriyle analiz edilmiştir. Sonuçlar, Inception-v3 modelinin en yüksek doğruluk (0,984) ve F1-skoruna (0,9263) ulaştığını, RegNetY-032'nin ise 0,941 AUROC değeriyle en güçlü ayırt edici performansı gösterdiğini ortaya koymuştur. Grad-CAM tabanlı açıklanabilirlik ve hata analizleriyle desteklenen çalışma, tekrarlanabilir ve klinik olarak yorumlanabilir bir değerlendirme çerçevesi sunmaktadır.

Anahtar Kelimeler

• Derin Öğrenme
• Medikal Tanı
• Tiroid Nodülü
• Ultrason Görüntüleme
• RegNetY-032

Classification of Thyroid Nodules from Ultrasound Images Using Deep Learning Methods

Öz

Thyroid cancer is one of the most prevalent malignancies of the endocrine system, and early diagnosis is critical for improving clinical outcomes. Although ultrasound imaging is the primary modality for thyroid nodule assessment, diagnostic interpretation remains highly dependent on expert experience. This study presents a standardized benchmarking framework for the systematic comparison of state-of-the-art deep learning architectures on thyroid ultrasound images. A total of 15 models, including RegNetY-032, ResNet50, EfficientNet, Inception, and YOLOv8 classification variants, were evaluated on the TN5000 dataset consisting of 5,000 images under identical preprocessing and training conditions. Performance was assessed using Accuracy, Precision, Recall, F1-score, and AUROC metrics. Experimental results showed that Inception-v3 achieved the highest diagnostic accuracy (0.984) and F1-score (0.9263), while RegNetY-032 demonstrated the strongest discriminative capability with an AUROC of 0.941. Supported by Grad-CAM-based explainability and error analyses, the proposed framework provides a reproducible and clinically interpretable evaluation of deep learning models for thyroid nodule classification.

Anahtar Kelimeler

• Deep Learning
• Medical Diagnosis
• Thyroid Nodule
• Ultrasound Imaging
• RegNetY-032

Kaynakça

1. Zhang X., Fan H., Yu Q., Tong X., Rao X., Tang T., Feng L., Shi Y., Xu Y., Wang M., Cheng Y. Global burden of thyroid cancer attributed to high body mass index and predictive trends: estimated results from the Global Health Data Study, 1990–2021. Thyroid Research, 18, 43, 2025.
2. De Carlos J., Garcia J., Basterra F. J., Pineda J. J., Ollero M. D., Toni M., Munarriz P., Anda E. Interobserver variability in thyroid ultrasound. Endocrine, 85, 730–736, 2024.
3. Schenke S. A., Solymosi T., Hegedüs L., Bonnema S. J. Considerable interobserver variation calls for unambiguous definitions of thyroid nodule ultrasound characteristics. European Thyroid Journal, 2024.
4. Loshchilov I., Hutter F. Decoupled weight decay regularization. Proceedings of the International Conference on Learning Representations (ICLR), 2019.
5. Vahdati S., Khosravi B., Robinson K. A., Rouzrokh P., Moassefi M., Akkus Z., Erickson B. J. A multi-view deep learning model for thyroid nodules detection and characterization in ultrasound imaging. Sensors, 11, 648, 2024.
6. Xu Y., Xu M., Geng Z., Liu J., Meng B. Thyroid nodule classification in ultrasound imaging using deep transfer learning. BMC Cancer, 25, 544, 2025.
7. Kim J., Kim M. H., Lim D. J., Lee H., Lee J. J., Kwon H. S., Kim M. K., Song K. H., Kim T. J., Jung S. L., Lee Y. O., Baek K. H. Deep learning technology for classification of thyroid nodules using multi-view ultrasound images. Endocrine Journal, 40, 216–224, 2025.
8. Chen C., Jiang Y., Yao J., Lai M., Liu Y., Jiang X., Ou D., Feng B., Zhou L., Xu J., Wu L., Zhou Y., Yue W., Dong F., Xu D. Deep learning to assist composition classification and thyroid solid nodule diagnosis: a multicenter diagnostic study. European Radiology, 34, 2323–2333, 2024.

9. Farooq M. U., Ghafoor H., Rehman A., Usman M., Chae D. K. GDSSA-Net: A gradually deeply supervised self-ensemble attention network for IoMT-integrated thyroid nodule segmentation. Internet of Things, 31, 101598, 2025.
10. Weng J., Tobriner W. B., Buda M., Yang J., Ho L. M., Allen B. C., Ehieli W. L., Miller C. M., Zhang J., Mazurowski M. A. Deep learning for classification of thyroid nodules on ultrasound: validation on an independent dataset. Radiology, 99, 60–66, 2023.
11. Göreke V. A. A novel deep-learning-based CADx architecture for classification of thyroid nodules using ultrasound images. Biomedical Signal Processing and Control, 15, 360–373, 2023.
12. Sun J., Wu B., Zhao T., Gao L., Xie K., Lin T., Sui J., Li X., Wu X., Ni X. Classification for thyroid nodule using ViT with contrastive learning in ultrasound images. Computer Methods and Programs in Biomedicine, 152, 106444, 2023.
13. Zheng Z., Su T., Wang Y., Weng Z., Chai J., Bu W., Xu J., Chen J. A novel ultrasound image diagnostic method for thyroid nodules. Diagnostics, 13, 1654, 2023.
14. Sharifi Y., Ashgzari M. D., Shafiei S., Zakavi S. R., Eslami S. Using deep learning for thyroid nodule risk stratification from ultrasound images. Heliyon, 3, 100082, 2025.
15. Chen Y. X., Hong H. Z., Gao Z. H., Hu Y. X., Yao L. Z., Liu J. T., Zhao Y., Cui G. W., Mao D. D., Zhao J. X. Global trend and disparity in the burden of thyroid cancer attributable to high body-mass index. BMC Public Health, 25, 1051, 2025.
16. Alyami J., Almutairi F. F., Aldoassary S., Albeshry A., Almontashri A., Abounassif M., Alamri M. Interobserver variability in ultrasound assessment of thyroid nodules. Medicine (Baltimore), 101, e31106, 2022.
17. Krönke M., Eilers C., Dimova D., Köhler M., Buschner G., Mirzojan L., Konstantinidou L., Makowski M., Nagarajah J., Navab N., Weber W., Wendler T. Tracked 3D ultrasound and deep neural network-based thyroid segmentation reduce interobserver variability. arXiv preprint, arXiv:2108.10118, 2021.
18. Haugen B. R., Alexander E. K., Bible K. C., et al. 2015 American Thyroid Association management guidelines. Thyroid, 26, 1–133, 2016.
19. Tessler F. N., Middleton W. D., Grant E. G., et al. ACR TI-RADS: White paper of the ACR committee. Journal of the American College of Radiology, 14, 587–595, 2017.
20. Liu T., Xie S., Yu J., et al. Artificial intelligence in thyroid nodule diagnosis: A review. Frontiers in Endocrinology, 14, 1126760, 2023.
21. Zhang, H., Liu, Q., Han, X., et al. TN5000: An ultrasound image dataset for thyroid nodule detection and classification. Scientific Data, 12, 1437, 2025.