RESNET-18 AND GRAD-CAM INTEGRATION FOR ENHANCED VISUALIZATION IN THYROID NODULE CLASSIFICATION

Thyroid nodules, characterized by abnormal thyroid cell growth, can indicate various health issues, including excessive iodine intake and inflammation.While most nodules are benign, the risk of malignancy increases annually.To alleviate the strain on healthcare providers and reduce unnecessary fine needle aspirations and surgeries, this study introduces a new deep learning framework aimed at accurately classifying thyroid nodules.We utilized a dataset of 508 ultrasound images from the Kaggle Repository, employing a pretrained ResNet18 model.Our method achieved impressive results: an average area under the curve (AUC) of 0.997, accuracy of 0.984, recall of 0.978, precision of 0.939, and an F1 score of 0.957.Additionally, we implemented Gradient-weighted Class Activation Mapping (Grad-CAM) to identify and analyze sensitive regions within ultrasound images, revealing significant differences in shape features between benign and malignant nodules.Overall, our model demonstrates the potential of deep learning in the accurate assessment of thyroid nodules using ultrasound imaging.