Representation Learning for Laparoscopic Imaging: A Study of Architectural Design and Multi-Level Knowledge Transfer

Semantic segmentation in laparoscopic surgery is critical for computer-assisted interventions, yet constrained by limited annotated data, dynamic environments, and significant class imbalance. This representation disparity results in suboptimal segmentation of smaller, often clinically significant, structures. This thesis systematically investigates deep learning strategies to address these challenges, focusing on architectural design, parameter sharing and transfer learning. We compare conventional shared-decoder frameworks with a novel multi-decoder architecture designed for organ-specific representational learning, evaluating various contemporary backbone architectures. Attention mechanisms prove particularly valuable in addressing underrepresented anatomical structures by enabling the network to focus on relevant spatial features. We explore transfer learning efficacy by pretraining on a large public laparoscopic dataset before fine-tuning on our target surgical dataset, analyzing how different encoder adaptation mechanisms interact with decoder design choices. Finally, we compare knowledge-sharing versus disjoint fine-tuning strategies for optimizing multi-class adaptation while preserving specialized representations. Our results demonstrate that the specialized multi-decoder approach consistently outperforms standard shared-decoder architectures across datasets, particularly improving segmentation for under-represented anatomical structures. Transfer learning substantially enhances performance metrics and accelerates convergence. Furthermore, preserving specialized features during fine-tuning proves superior to knowledge-sharing approaches, highlighting the importance of maintaining structure-specific representations. This research establishes that specialized multi-decoder architectures, combined with effective transfer learning strategies and attention mechanisms, significantly advance multi-organ segmentation in laparoscopy by promoting robust structure-specific representational learning and mitigating persistent representation bias.