Accurately distinguishing between laryngeal diseases is a difficult undertaking that frequently leads to diagnostic mistakes due to expert subjectivity. Most existing automated systems are limited to a binary classification of benign versus malignant tissues. This study introduces a novel deep learning framework to overcome this limitation by performing a fine-grained classification of Contact Endoscopy - Narrow Band Imaging (CE-NBI) into eight distinct classes, including Squamous Cell Carcinoma (SCC), various grades of dysplasia, and papillomatosis. In the current study a U-shaped network architecture augmented with a specialized attention block module is proposed. This mechanism empowers the network to selectively focus on the most discriminative features within the endoscopic images. The proposed approach outperforms state-of-the-art techniques, according to comparative trials that support the methodology. Notably, its efficient design makes it highly compatible with low-memory hardware, presenting a practical and effective solution for automated laryngeal cancer diagnosis.