Head and neck cancers pose significant diagnostic challenges due to their complex anatomy, high recurrence rates, and subtle early symptoms. This study utilizes Convolutional Neural Networks (CNNs) as the base algorithm to develop two distinct models: one using U-Net for tumor segmentation and the other using ResNet for classification. The U-Net model, leveraging an encoder-decoder structure with skip connections, achieved 97% accuracy, 96% precision, 89% recall, and an F1-score of 92%, demonstrating its effectiveness in segmenting tumor regions. Meanwhile, the ResNet-based model, designed for robust feature extraction and classification, attained 87% accuracy, 88% precision, 87.7% recall, and an F1-score of 87.4%, confirming its reliability in distinguishing cancerous and non-cancerous regions. Both models were trained on expert-annotated PET scan datasets and evaluated using precision, recall, and segmentation accuracy. A comparative analysis of these models provides insights into their performance and potential integration, supporting radiologists in improving diagnostic workflows and enabling early cancer detection.