Better Together Than Separate: Prediagnosis of Thyroid Nodules Through Ultrasound Imaging Using CNN, ViT and Hybrid Models

The increase in thyroid cancer cases registered from 1990 to the last decade by 20 % worldwide, has created a need for faster and more accurate diagnostic tools for thyroid nodules. Deep Learning architectures, including Convolutional Neural Networks and Vision Transformers, have been developed to assist in diagnosing whether nodules are benign or malignant. Data Augmentation techniques such as DCGAN were also utilized. Hybrid models combining these type of architectures were also trained. The hybrid model developed by Google, re-trained with the TNCD dataset, achieved the best results with values of 77.20% for Accuracy, 77.97% for Recall, and 67.65% for Precision. The success of the hybrid approach depends on the architectures combined and whether they have been pre-trained. These findings suggest that a pre-trained model combining CNN and ViT is superior to using them independently, highlighting the potential of combining these architectures for improved the pre-diagnostic accuracy.