Transfer Learning in Thyroid Malignancy Prediction Explained with Grad-CAM

Thyroid cancer is a significant global health concern requiring precise and prompt interventions. Diagnosing thyroid cancer, particularly through ultrasound imaging, is challenging due to the need for both high accuracy and explainability in clinical settings. Existing AI-based methods for classifying thyroid nodules have limitations in accuracy. This paper proposes a novel AI approach to enhance the accuracy of thyroid cancer classification using deep neural networks. We explored the application of transfer learning techniques. Pre-trained models (VGG-16, DenseNet-201, Inception-v3) are employed with additional custom dense classification layers for classifying and predicting thyroid cancer from ultrasound images. The convolutional layers of the pre-trained models are selectively fine-tuned and the performance is monitored. VGG-16 model with fine-tuning of block 5 convolutional layers consistently showed good performance. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) is used to explain the models’ decision-making process. The proposed approach achieved state-of-the-art performance with accuracies of 94.97% (VGG-16), 90.48% (DenseNet-201), and 90.23% (Inception-v3). Furthermore, Grad-CAM generates heatmaps highlighting crucial features in the ultrasound images for predicting cancer. Grad-CAM goes beyond simply identifying discriminative regions. It delves deeper by analyzing gradients to pinpoint the most critical image features that influence the model’s prediction. This enhanced explainability fosters trust and collaboration between doctors and AI models in the context of medical diagnosis. This research integrates medical expertise and explainability into the proposed AI model, aiming to improve the accuracy and practicality of thyroid cancer diagnosis for doctors, ultimately benefiting patients.