Data-Efficient Deep Learning for Surgical Perception in Minimally Invasive Surgery

Minimally invasive surgery has become the standard of care across a wide range of clinical specialities, with endoscopic and robotic procedures offering shorter recovery times, reduced complication rates, and superior ergonomics compared to open surgery. However, both paradigms rely on a monocular two-dimensional video feed as the sole visual channel into the surgical site, resulting in diminished depth perception, blind spots, and an intrinsic loss of spatial context that make tissue examination, diagnosis, and instrument manipulation substantially more demanding. Artificial intelligence, and deep learning in particular, has emerged as a transformative route to restore this missing perceptual information and to enable higher levels of surgical scene understanding. Yet its clinical deployment is hindered by two compounding barriers: the scarcity of labeled medical data and the strict real-time inference requirements of the operating room. The first contribution of this thesis addresses the recovery of three-dimensional geometry from a single endoscopic camera. We propose \textit{BodySLAM}, a generalized monocular visual SLAM framework that integrates three components: \textit{CycleVO}, a novel unsupervised pose estimation module that learns robust camera motion representations from unlabeled endoscopic video through a GAN-based image-to-image translation pretext task; the Zoe architecture for monocular depth estimation, leveraged without fine-tuning to exploit its strong cross-domain generalization; and a three-dimensional reconstruction module coupled with an Unscented Kalman Filter and pose-graph optimization for scale-consistent mapping. Evaluated on the three largest public endoscopic MVSLAM benchmarks across laparoscopy, gastroscopy, and colonoscopy, BodySLAM matches or outperforms four state-of-the-art methods, while CycleVO attains the lowest inference time among all competitors, simultaneously addressing the data-scarcity and real-time barriers. The second contribution targets the downstream perception task of classification under extreme label scarcity, a regime in which deep networks traditionally fail to converge. We introduce \textit{SPARSE}, a three-player GAN-based semi-supervised learning framework that alternates supervised and unsupervised phases, leverages class-conditioned image-to-image translation to enrich feature representations from real unlabeled images rather than generating them from noise, and employs a confidence-weighted temporal ensemble for reliable pseudo-labeling. Across eleven MedMNIST benchmarks spanning diverse imaging modalities, SPARSE achieves statistically significant improvements over six state-of-the-art semi-supervised methods with as few as five labeled samples per class, demonstrating that robust classification is attainable under data conditions that mirror the realities of clinical medical image annotation. Collectively, the contributions of this thesis address the compounding challenges of data scarcity and real-time inference that currently stand between deep learning and its clinical deployment in endoscopic and robotic surgery. By unifying unsupervised pose estimation, generalizable depth prediction, and image-translation-based semi-supervised classification into coherent, computationally efficient pipelines, the proposed approaches pave the way toward AI-enhanced surgical perception systems that are simultaneously accurate, data-efficient, and viable for the operating room.