Background: Thyroid cancer presents a significant global health challenge, necessitating precise and timely interventions. Several recent studies have used transfer learning techniques for thyroid cancer classification from ultrasound images. While existing research has achieved promising results using transfer learning for thyroid cancer classification from ultrasound images, two key limitations remain unaddressed: Use of domain knowledge, and Explainability. Existing state-of-the-art deep learning models fail to leverage domain knowledge effectively. Aim: This work addresses this gap by proposing a novel explainable deep learning architecture incorporating domain knowledge. The proposed model incorporates feature extraction by VGG-16 convolutional layers, a domain-specific knowledge block, and a classification network for accurate malignancy classification. A Gradient-weighted Class Activation Mapping (Grad-CAM), integrated with our model, is an explainability tool. Method: The proposed customised knowledge-guided block is tailored based on the importance of high-ranked features such as composition, echogenicity, shape, margins, and calcifications. The 18-layer architecture includes 13 convolutional layers of the VGG-16 base model for feature extraction, 1 layer in domain knowledge-guided block, 1 concatenation and flattening layer, and 3 dense classification layers for accurate classification of thyroid ultrasound images. Results: Our approach achieves state-of-the-art predictive performance, with an accuracy of 92.84%. Grad-CAM generates heatmaps highlighting regions and features in ultrasound images crucial for cancer classification. By incorporating domain knowledge into deep learning architecture, we enable its effective application in real-world clinical settings, specifically for the classification of thyroid cancer. Our approach achieves state-of-the-art accuracy, while Grad-CAM empowers clinicians to understand the model’s reasoning.