A Controlled Comparison of U-Net Variants for Surgical Instrument Segmentation in Minimally Invasive Surgery

Background: Surgical instrument segmentation is a relevant computer vision task in minimally invasive surgery, where binary masks can be used to separate surgical instruments from the background. This study compares selected U-Net architectural configurations for binary surgical instrument segmentation on the EndoVis 2020 dataset. Methods: The input images were cropped to 538×701 pixels and paired with binary segmentation masks. To reduce memory usage, the models used an internal resizing step to process images at 256×256 resolution, while predictions were resized back to the original image size for evaluation. Several U-Net variants were implemented and trained under the same conditions, including changes in network depth, filter configuration, kernel size, and pooling strategy. Results: The baseline U-Net achieved an IoU of 0.97 and a Dice score of 0.98. Among the successfully trained configurations, the increased-depth variant achieved the highest numerical Dice score, with an IoU of 0.97 and a Dice score of 0.99. The increased-pooling and reduced-depth variants produced lower IoU and Dice scores within the present experimental setup, while the variable-kernel configuration produced results close to those of the baseline model. The increased-filter and reversed-filter configurations did not produce valid segmentation outputs under the common training protocol. Conclusions: Within the present experimental setting, the extended-depth configuration achieved the highest numerical Dice score among the successfully trained variants. The results are interpreted as a controlled comparison of U-Net configurations for binary surgical instrument segmentation rather than as evidence of general architectural superiority. Because the evaluation used a frame-level split, frames originating from the same surgical recording may occur in different subsets. Although the two-dimensional models process each frame independently without temporal or sequence-level inputs, complete sequence-level statistical independence is not guaranteed. The reported differences should therefore be interpreted within the applied experimental protocol. Future work will include sequence-level evaluation, repeated experiments, and testing on additional datasets.