The diagnosis of vocal fold (VF) lesions relies on examinations such as laryngoscopy and voice analysis, which are highly dependent on clinicians' experience. This leads to a relatively higher risk of misdiagnosis and missed diagnosis among junior physicians. In recent years, with the rapid advancement of artificial intelligence (AI), numerous deep learning (DL)-based methods have emerged in this field. Early research primarily focused on single-modality image analysis, such as classifying white light or narrow-band images into benign or malignant lesions using convolutional neural networks (CNNs). However, such methods often fail to fully integrate complementary information from different modalities and fall to meet the clinical demands for multi-classification and risk stratification. Recently, DL and multimodal fusion have gradually become research hotspots. It enables the extraction of complementary multi-category feature information by integrating laryngoscopic images, videos, voice, and clinical text data (e.g., laryngoscopy reports and medical record information), to construct an end-to-end intelligent diagnostic system. This narrative review summarizes the research progress of DL and multimodal fusion in the diagnosis, classification, and severity grading of VF lesions over the past five years (2020–2025). Studies demonstrate that multimodal DL models outperform single-modality models across multiple tasks, which significantly improves the identification and classification accuracy of VF lesions. These models exhibit promising performance. However, DL and multimodal fusion still face numerous challenges, and their clinical translation remains difficult.