Vulvar cancer has significant implications for women's health and quality of life, emphasizing the need for prompt diagnosis and treatment. Surgical treatment aims to remove both visible and nonvisible malignant cells, incorporating a safety margin beyond the tumour perimeter. However, accurately assessing cancer spread and determining the optimal safety margin size remains challenging. Hyperspectral imaging allows non-invasive identification of vulvar tumour boundaries, capturing both spatial and spectral details of the region of interest. However, tissue type classification in hyperspectral imaging is complex due to the abundance of wavebands and limited training samples. Artificial intelligence (AI) methods show promise for accurate and automated tumour classification in hyperspectral imaging. In this study, four AI models (Support Vector Mechanism, Neural Network, 1D and 3D Convolutional Neural Networks) were trained to classify tumour tissue, skin tissue, and mucosa tissue. Additionally, the study aimed to explore the use of specific explanatory parameters (Tissue Water Index, Near Infrared Perfusion Index, and Tissue Haemoglobin Index) to reduce the spectral dimension, potentially improving acquisition time and interpretability without compromising classification accuracy. The study included 25 patients (mean age 71), with hyperspectral cubes captured from each patient. All cubes contained healthy skin tissue, while 22 cubes included tumour tissue and 10 cubes contained mucosa tissue. All models exhibited comparable performance for tumour detection, with F1 scores ranging between 0.91 and 0.93 and AUC-ROC scores ranging between 0.91 and 0.94. Moreover, the Neural Network trained only on the explanatory parameters achieved excellent results for tumour detection with an F1 score of 0.93 and AUC-ROC of 0.91. Overall, combining hyperspectral imaging with all models shows great potential for in vivo tumour detection, while leveraging physiological explanatory parameters with a Neural Network model can enhance acquisition time and data interpretability without compromising the classification accuracy.