Gesture recognition has emerged as a promising technology in the field of surgical support, offering a novel means of enhancing communication and control during the preoperative phase and within the operating room. Surgeons and medical professionals are increasingly adopting gesture recognition systems to streamline their workflows and reduce contamination risks associated with traditional input methods; moreover, this interaction method is widely adopted for Augmented Reality devices such as the Head-Mounted Display (HMD) Hololens 2. This thesis aims to develop and evaluate various neural network models to determine the most effective one for supporting surgery planning and execution. By implementing models that take images and landmarks as inputs, the goal is to find the model with the best compromise between frame rate and accuracy.