Background: The application of Deep Learning (DL) in Minimally Invasive Gynecological Surgery (MIGS) has accelerated rapidly, driven by the need for enhanced intra-operative navigation, surgical phase recognition, and predictive analytics. The efficacy of these applications is fundamentally determined by the underlying neural network architectures, yet a comprehensive technical synthesis of their comparative performance and clinical utility in the gynecological context is absent from the literature.

Objective: To provide a comprehensive scoping review of the DL architectures deployed in MIGS, evaluating their technical mechanisms, clinical applications, performance metrics, and limitations.

Methods: A scoping review was conducted following PRISMA-ScR guidelines, drawing from a systematic search of 55 included studies identified across five major databases. Studies explicitly detailing neural network architectures applied to gynecological surgical tasks were analyzed, with a focus on Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Deep Neural Networks (DNNs), Transformer architectures, and hybrid models.

Results: CNNs remain the foundational architecture for spatial feature extraction in surgical video analysis, achieving accuracy rates of 85–97% in anatomical landmark identification and instrument detection. LSTM networks, frequently integrated with CNNs in hybrid architectures, significantly improve surgical phase recognition by capturing temporal dependencies across video sequences. Transformer architectures, including Vision Transformers (ViTs), have demonstrated superior performance in complex multi-modal tasks, outperforming CNN-LSTM models in surgical workflow understanding. Emerging applications include augmented reality navigation, automated endometriosis lesion recognition, and AI-driven prediction of perioperative outcomes.

Conclusion: The DL landscape in MIGS is evolving from static image analysis toward dynamic, spatio-temporal surgical understanding. While CNNs and LSTMs constitute the current standard, Transformers represent the emerging frontier. Critical priorities for future development include model interpretability, real-time edge deployment, and standardized benchmarking across diverse gynecological datasets.