Enhancing Thyroid Cancer Detection in Ultrasound Images Using Augmentation-Driven Deep Learning Framework

Thyroid cancer has become the most rapidly increasing endocrine malignancy, and its incidence is on the rise globally. Deep learning methods demonstrated reliable prospects in improving thyroid cancer diagnosis based on ultrasound imaging. But data scarcity and class imbalance make it challenging to build robust models. In this paper, we propose an image augmentation approach for enhancing thyroid cancer classification and segmentation using deep learning. The Thyroid Ultrasound-Image Database, which contains 2,450 images with five ACR-TIRADS levels, was applied. Pre-processing: Noise reduction, inverse intensity, and normalization were performed. Geometric transformations, photometric augmentations, GAN-based synthetic image generation and domain-specific augmentations were utilized. For classification and segmentation tasks, CNN-based architectures (e.g., VGG, ResNet) as well as transformer-based models were employed, along with U-Net variants. Conclusions: Data augmentation was found to lead to a substantial improvement in the models' ability to generalize, with 10–15% gains in accuracy. Various augmentation techniques were performed differently, and the combination of multiple techniques was the most accurate. Qualitative results indicated the robustness of feature extraction, and quantitative comparisons showed that our method was competitive with several state-of-the-art methods. The present approach has potential clinical applications to help radiologists in the early detection of thyroid cancer. Potential future work involves adapting advanced augmentation methods and multimodal fusion to further increase the classification performance.