While thyroid nodules are frequently seen and can be detected in imaging and autopsy tests up to 65% of the time, only roughly 10% of them are cancerous. This emphasizes how crucial precise preoperative classification is to guaranteeing suitable clinical judgment and preventing overtreatment. In this study, we propose a deep learning-based diagnostic framework that integrates Visual Geometry Group (VGG)-19 with Gradient-weighted Class Activation Mapping (Grad-CAM) for tumor localization in thyroid ultrasound images. The VGG-19 model, pre-trained and fine-tuned for domain-specific learning, is leveraged for robust feature extraction. At the same time, Grad-CAM enhances interpretability by producing heatmaps that visually highlight areas influencing the model’s classification. We evaluated our system on publicly available datasets—Hebei (508 images) and DDTI (591 images)—and benchmarked its performance against state-of-the-art architectures like ResNet18, DenseNet, and EfficientNet. The combination of VGG-19 and Grad-CAM ensemble achieved 97.35% accuracy, 95.75% sensitivity, and 98.4% specificity, outperforming traditional models and offering clinical-grade reliability. Grad-CAM visualizations accurately highlight suspicious regions, offering radiologists valuable second-opinion insights. This enhances trust in AI diagnoses while ensuring high accuracy and interpretability for practical use in thyroid cancer screening.