SegSTRONG-C

Challenge Description

Accurate segmentation of tools in robot-assisted surgery is a foundational aspect of machine perception, facilitating various downstream tasks, including augmented reality feedback. While existing feed-forward network-based methods exhibit excellent performance in the absence of corruption in test cases, the susceptibility to even minor corruptions can significantly impair the model's performance due to the inherent overfitting nature of such networks. This vulnerability becomes particularly consequential in surgical applications, where high-stakes decisions are commonplace. Prior efforts, such as benchmarking methods on ImageNet-C for image classification, have explored the robustness of models by introducing artificial noise to test images. The CaRTS approach introduces a novel pipeline designed to achieve robust segmentation of robot tools against realistic corruptions. To assess algorithm robustness against non-adversarial corruptions, we expand the dataset in CaRTS to address the challenge of robust robot tool segmentation against non-adversarial corruption. Our goal is to encourage algorithms to exhibit robustness to unforeseen yet plausible complications that may arise during surgery, such as smoke, over-bleeding, and low brightness. The training set comprises 14 mock endoscopic surgery sequences without corruption and corresponding binary segmentation masks where "1" represents robot tools and "0" represents tissue background. In the testing set, there are three sequences with non-adversarial corruptions (smoke, over-bleeding, and low brightness) and digital corruption (ImageNet-C). Additionally, three sequences with an alternative background serve as validation. Participants are challenged to train their algorithms solely on uncorrupted sequences while achieving high performance on corrupted ones for the binary robot tool segmentation task. Successfully achieving this non-adversarial robustness in this benchmark is paramount for the translation of research algorithms into real-world applications. It ensures that these algorithms can navigate and perform effectively in the face of unexpected but reasonable complexities encountered during surgical procedures.

Keywords: Non-adversarial Robustness, Surgical Tool Segmentation, Robotics Surgery, Minimally Invasive Surgery, EndoVis