Thyroid nodule detection is critical for early diagnosis and treatment of thyroid-related diseases. This study presents an efficient lightweight CNN-based approach for thyroid nodule detection in CT scan images, addressing resource constraints. We used a dataset from the Figshare repository, consisting of 150 healthy CT scans and 134 nodule images. Two lightweight CNN models, MobileNetV2 and ShuffleNet, were employed in transfer learning approaches. MobileNetV2 achieved an accuracy of 85.19%, sensitivity of 100%, specificity of 70.37%, and an F1 score of 87.1%, while ShuffleNet achieved an accuracy of 77.78%, sensitivity of 59.26%, specificity of 96.3%, and an F1 score of 72.73%. To enhance performance, we implemented a feature fusion technique, combining deep features from MobileNetV2 and ShuffleNet, which were then classified using an SVM model. This fusion approach significantly improved the accuracy to 97.98%, with an AUC of 0.978 for both normal and nodule cases. The proposed framework demonstrates potential for accurate and efficient thyroid nodule detection in resource-constrained environments.