A Dataset and Benchmark for Robot-Assisted Radical Prostatectomy With Lymphadenectomy in Surgical Workflow Undertstanding

Objective: Surgical workflow understanding, especially surgical phase and step recognition, plays a crucial role for improving context-aware computer-assisted system in modern operating room. However, many state-of-the-art methods are limited by the scarcity of diverse and well-annotated datasets and have only been experimented with laparoscopic and non-robotic datasets, which makes the advancement of intelligent systems and model generalization validation in varied and intricate real-word surgical situation extremely challenging. Methods: To address this challenge, we introduce a video dataset with expert-annotated workflow information for Robot-assisted Radical Prostatectomy with Lymphadenectomy. We record surgical videos performed on 10 patients using the da Vinci Xi robot at the European Institute of Oncology in Milan, Italy. The videos are annotated under the guidance of experienced surgeons, with the entire dataset exceeding 25.3 hours. Subsequently, we perform a comprehensive evaluation using CNN-based and Transformer-based methods to establish potential benchmarks for surgical phase and step recognition on the dataset. Each model is rigorously assessed using five-fold cross-validation to ensure reliability and general- izability. Results: Eight representative models, including six CNN-based (SAHC, PhaseNet, SV-RCNet, TMR-Net, TeCNO, and Trans-SVNet) and two Transformer-based approaches (Timesformer and Surgformer), are evaluated across both tasks. For surgical phase recognition, Surgformer achieves the highest performance with the accuracy of 90.3%. For surgical step recognition, Surgformer_16_4 yields comparable results to that of Timesformer_16_4, around 78%. Significance: We present a complete procedural dataset in the data-scarce domain of RARPL, enabling robust benchmarking for surgical workflow recognition and bridging the gap between algorithm development and clinical practice. The dataset is accessible at the link: https://zenodo.org/records/15228400.