This work presents an innovative approach to advancing computer-assisted surgical (CAS) systems by addressing challenges in training data quality and realism.
In this study, the authors introduce a comprehensive pipeline for creating high-quality synthetic data tailored for modern CAS systems. This pipeline integrates an advanced surgical simulator capable of generating complex annotations that surpass those found in existing public datasets. The simulator also models intricate surgical interactions, including the dynamics between instruments and deformable anatomical structures, ensuring a more realistic simulation environment.
To further enhance data realism, the researchers developed a novel image-to-image translation method based on Stable Diffusion (SD) and Low-Rank Adaptation (LoRA). This technique minimizes the visual gap between synthetic and real-world images while preserving the simulator’s detailed annotations. By leveraging only a small amount of real-world data, the method enables efficient training and generalizes well to practical applications, thereby improving CAS training and guidance.
This innovative pipeline has been validated experimentally and is a significant step forward in bridging the gap between synthetic and real-world datasets for surgical applications. The dataset and code are available for the research community github.com/SanoScience/SimuScope
Authors: Sabina Martyniak, Joanna Kaleta, Diego Dall’Alba, Michał Naskręt, Szymon Płotka, and Przemysław Korzeniowski
DOI: 10.48550/arXiv.2412.02332
READ HERE