Laparoscopic surgery is one of the most promising methods of traumatic endos-copy that can be automated. The current level of medical robotics allows for surgical interventions to be performed only under the surgeon’s complete control. Partial automation of this process requires solving many problems. An essential component of this process is providing automated operating space analysis. This is what makes feedback possible, which serves to assess control, positioning, manipulation, etc. One of the peculiarities of laparoscopic image analysis is the partial high variability of the overall surgical field, which depends on many factors. The article describes the process of detecting and segmenting the operating scene, namely the surgical instrument, using convolutional neural networks. The construction of neural network architectures, dataset organization, and training were carried out using the Detectron2 software platform. The practical results obtained indicate the high accuracy of the corresponding models. The results were obtained for different neural net-work architectures. As a result of the study, an initial version of the dataset on surgical instrument segmentation in laparoscopic images was obtained. The direction of further research is to improve detection and segmentation results by expanding the data sets (training set and test set).