A Vision Transformers Approach for Surgical Monitoring with Algorithmic Framework and Experimental Evaluation

IoMT-enabled surgical monitoring systems rely on AI for real-time decision making. Although significant progress has been made, problems such as tissue variability, dataset imbalance, and segmentation errors prevail. This therefore necessitates innovative solutions to ensure more safety, accuracy, and efficiency in robotic surgeries. We propose to design an IoMT-enabled framework for surgical monitoring systems using Vision Transformers, super-resolution imaging, and clustering algorithms for safety, image clarity enhancement, and balanced data representation. It incorporates dynamic attention-based keyframe extraction, super-resolution imaging with PWLC-SRGAN, segmentation using U-Net++, and classification with Vision Transformers, applying entropy regularization. The algorithms handle tissue variability and enhance safety by providing real-time alerts. The framework had a Dice score of 0.94, with classification accuracy at 98.9% and a sensitivity of 97.2%. Real-time notifications detected tool misplacements in milliseconds, thus enhancing procedural safety. False positives decreased to 0.5%, surpassing traditional approaches in all parameters. The framework transforms IoMT-enabled surgical monitoring with high precision, robust safety safeguards, and real-time feedback. Institutional Review Board approval for the study was granted under IRB-2023-MED-447 at REVA University Hospital. Each participant provided written informed consent.