<p>Segmentation of endoscopic surgical instruments can assist surgeons in making decisions to ensure surgical safety. Semantic segmentation based on deep learning relies heavily on large-scale annotated data, which is both time-consuming and labor-intensive. Fortunately, semi-supervised learning offers a means to alleviate this challenge. However, Existing semi-supervised segmentation methods are rarely specifically designed for endoscopic surgical instruments. To address the problem of insufficient presence of endoscopic surgical instruments within the visual field, leading to bias in model learning, we devise an image incremental augmentation strategy aimed at augmenting the presence of endoscopic surgical instruments within the images. Furthermore, to enhance the model’s focus on inconsistent regions within segmented targets, we devise an image decrement augmentation strategy aimed at improving the segmentation accuracy of endoscopic surgical instruments’ inconsistent regions. The adopted network has two different decoders and computes the consistency of their outputs. Several consistency losses are devised to reinforce supervision among networks employing different data augmentation strategies in their outputs. Extensive experiments on both a private dataset and a public dataset corroborates the effectiveness of the proposed method. Our method achieves a Dice similarity coefficient (DSC) of 92.96% and 93.36% on the EndoVis2017 dataset with 5% and 10% labeled data, respectively, outperforming state-of-the-art methods. On the private dataset, our method achieves a DSC of 97.64% and 97.78% with 5% and 10% labeled data, respectively, further demonstrating its superiority. Ablation studies confirm the effectiveness of the proposed data augmentation strategies, with incremental and decremental strategies contributing to a 0.72% and 0.41% improvement in DSC, respectively.</p>