Early detection of thyroid cancer is essential because it enables high cure rates and often allows for minimally invasive treatments, significantly improving outcomes. When found early, thyroid cancer is typically localized, reducing the risk of metastasis and the need for aggressive interventions. Early detection also enhances quality of life by lowering the chances of complications, such as vocal cord damage or calcium imbalance, that can arise from advanced treatments. Early cancer detection can also result in more affordable care by reducing the need for needless treatments and providing options like active surveillance for some low-risk cases. In this work, we present a deep convolutional neural network (Deep-CNN) model as a novel method for detecting thyroid cancer stages. This approach combines preprocessing, segmentation, and feature extraction techniques to effectively classify Thyroid Imaging Reporting and Data System (TI-RADS) stages. It utilizes robust Principal Component Analysis (PCA) and the effective Gray Level Co-occurrence Matrix (GLCM), achieving a remarkable accuracy of 89.2% across TI-RADS stages 1 to 5, outperforming existing models. PCA in image processing reduces data dimensionality by retaining essential features, aiding in compression, noise reduction, and feature extraction. This enhances efficiency in applications like face recognition and object detection whereas A Gray Level Co-occurrence Matrix (GLCM) analyzes image texture by mapping spatial relationships between pixel intensities. It reveals texture features like contrast and homogeneity, aiding applications in medical imaging, remote sensing, and pattern recognition. Unlike previous research, which primarily examined the margins of thyroid nodules, this work focuses on detecting calcium flecks, filling a critical gap in the field. The findings offer valuable potential for advancing thyroid cancer detection and patient care.