Surgical instrument segmentation in images is crucial for improving precision and efficiency in surgery, but it currently relies on costly and labor-intensive manual annotations. An unsupervised approach is a promising solution to this challenge. This paper introduces a surgical instrument segmentation method using unsupervised machine learning, based on the K-means algorithm, to identify Regions of Interest (ROI) in images and create the image ground truth for neural network training. The Gamma correction adjusts image brightness and enhances the identification of areas containing surgical instruments. The K-means algorithm clusters similar pixels and detects ROIs despite changes in illumination, yielding an efficient segmentation despite variations in image illumination and obstructing objects. Therefore, the neural network generalizes the image features learning for instrument segmentation in different tasks. Experimental results using the JIGSAWS and EndoVis databases demonstrate the method's effectiveness and robustness, with a minimal error (0.0297) and high accuracy (0.9602). These results underscore the precision of surgical instrument detection and segmentation, which is crucial for automating instrument detection in surgical procedures without pre-labeled datasets. Furthermore, this technique could be applied in surgical applications such as surgeon skills assessment and robot motion planning, where precise instrument detection is indispensable.

Accurate needle insertion during epidural anesthesia is challenging due to strong dependence on clinician experience and the limited integration of guidance modalities that simultaneously provide visual feedback and physical motion constraints. Current approaches, including ultrasound guidance and augmented reality visualization, mainly offer passive assistance and do not actively regulate insertion trajectory and depth, which may lead to variability in accuracy and increased risk of complications. This work presents a multimodal human–machine assistance system that combines augmented reality guidance with virtual fixtures to support lumbar epidural needle placement. A Tuohy needle is coupled to a haptic device interacting with a patient-specific L3–L4 lumbar phantom fabricated using 3D printing and ballistic gel. A model-based force profile reproduces the mechanical response of anatomical layers during insertion. Three experimental conditions are evaluated: freehand execution, augmented reality guidance with trajectory and depth visualization, and cooperative guidance using virtual fixtures defined by a cylindrical corridor and a depth-limiting plane. Results show a progressive reduction in mean depth error from 6.82 ± 3.46 mm (freehand) to 4.96 ± 2.41 mm (augmented reality) and 2.21 ± 1.73 mm (virtual fixtures). These findings indicate that the integration of visual and haptic guidance significantly enhances insertion precision and control. The proposed approach highlights the potential of multimodal human–machine cooperation for safer training and assisted interventions.