Enhancing Self-Supervised Monocular Depth Estimation in Endoscopy via Feature-Based Perceptual Loss

Abstract

In recent years, self-supervised learning methods for monocular depth estimation have garnered significant attention due to their ability to learn from large amounts of unlabelled data. In this study, we propose further improvements for endoscopic scenes based on existing self-supervised monocular depth estimation methods. The previous method introduce an appearance flow to address brightness inconsistencies caused by lighting changes and uses a unified self-supervised framework to estimate both depth and camera motion simultaneously. However, to further enhance the model's supervisory signals, we introduce a new feature-based perceptual loss. This module utilizes a pre-trained encoder to extract features from both the synthesized and target frames and calculates their cosine dissimilarity as an additional source of supervision. In this way, we aim to improve the model's robustness in handling complex lighting and surface reflection conditions in endoscopic scenes. We compare the performance of using two pre-trained CNN-based models and four foundational models as encoder. Experimental results show that our improve method further enhances the accuracy of depth estimation in medical imaging. Additionally, it demonstrates that features extracted by CNN-based models, which are sensitive to local details, outperform foundation models. This suggests that encoders for extracting medical image features may not require extensive pre-training, and relatively simple traditional convolutional neural networks can suffice.