Gaussian Process Prior Variational Autoencoder for Endoscopic Videos

Endoscopic video analysis is frequently hindered by transient artifacts such as specular reflections, motion blur, and missing frames. These issues disrupt clinical interpretation and downstream tasks like 3D reconstruction. This paper investigates whether explicitly modeling the temporal continuity of endoscopic video sequences—rather than treating frames as independent samples—can improve image restoration and provide meaningful uncertainty estimates for clinicians.

The authors introduce a Gaussian Process Prior Variational Autoencoder (GPVAE) framework. Unlike standard VAEs that assume independent latent variables, the GPVAE imposes a Gaussian process (GP) prior over the latent space. This forces the model to learn a continuous, smooth latent trajectory over time. To make this computationally feasible for long endoscopic videos, the authors employ two scalable GP approximations: the Hierarchical Prior Approximation (HPA) and the Sparse Precision Approximation (SPA). The framework is further enhanced by integrating endoscopy-specific encoders, such as the EndoVAE backbone and pretrained Vision Transformer (ViT) encoders from GastroNet-5M, and a DUCKNet-based masking pipeline to ignore corrupted pixels during training.

The GPVAE framework significantly outperforms standard VAE baselines on the C3VDv2 colonoscopy dataset. Specifically, the best GPVAE variants achieved an average reduction in image reconstruction RMSE of 21.9% (up to 26.1%). Furthermore, the model improved downstream trajectory estimation (visual odometry and PoseNet) by 12.7% on average. A key advantage of this probabilistic approach is the generation of per-frame uncertainty estimates, which provide a confidence signal for the restored images, helping to identify regions or frames that may require additional clinical review.