Machine Learning in Medical Robotics of Diagnostics, Surgery and Rehabilitation

Recently, the integration of Machine Learning (ML) into medical robotics has revolutionized the fields of diagnosis, surgery and rehabilitation by enabling minimally invasive procedures and improving recovery outcomes and transforming them into assistants capable of interpreting complex data and supporting real-time clinical decisions. This review mainly focuses on ML applications in three areas: diagnostics, surgery, and rehabilitation. In diagnostics, attention-based U-Net variants like OP-U-Net uses optical flow and channel attention for real-time vascular segmentation in robotic ultrasound, and there is also skin cancer detection by using Hybrid Convolutional Neural Network (CNN) frameworks, while voice analysis using supervised ML models enables early Parkinson diagnosis. In surgery, deep learning enables autonomous vascular access and semantic segmentation of instruments by using architecture such as TernausNet and LinkNet. Super-resolution in endomicroscopy is achieved through synthetic data training, overcoming hardware constraints and improving intraoperative imaging. In rehabilitation, reinforcement learning frameworks like Flexible Policy Iteration (FPI) enhance robotic knee control by integrating experience replay and prior knowledge. Additionally, ML-driven wearable systems for stroke rehabilitation enable real-time gesture recognition and robotic hand assistance, supporting at-home recovery. These methods bring substantial improvements in accuracy, adaptability, and clinical relevance. However, challenges in generalizability, interpretability, and data privacy remain. Solving these issues by explainable AI, federated learning, and domain-informed architectures is crucial for the future integration of intelligent robots into everyday healthcare.