A benchmark of methods for surgical instrument segmentation in endoscopic pituitary surgery

PURPOSE: The development of impactful AI solutions for minimally invasive surgeries is reliant on fundamental tasks such as surgical instrument segmentation. The fine-grained information provided can support downstream tasks such as skill assessment, intra-operative decision support, and outcome prediction. However, many procedures which stand to benefit significantly from these applications are under-represented in literature as a result of the heavy requirements for data collection and annotation. METHODS: This study examines an in-house labelled surgical instrument dataset for endoscopic transsphenoidal pituitary surgery and compares the performance of a number of established segmentation models, with both transformer-based and convolutional architectures. RESULTS: Our results show that the highest performance was achieved by the EoMT model, with a DSC of 0.7420 while maintaining a more balanced performance across minority classes. Additionally, our results show the challenges caused by the use of real surgical footage, with highly variable frame contents, and significant class imbalances. This is reflected in the reduced performance across classes with low frequencies which are similar in appearance to more popular classes. CONCLUSION: These findings highlight model architectures with potential for development and application to surgical instrument segmentation in pituitary surgery and additionally emphasise the need for larger datasets and further work to include under-represented surgeries with novel challenges.