Soft tissue surface deformation detection in endoscopic images

Soft tissue surface deformation in endoscopic images breaks the alignment between preoperative models and the intraoperative view, which affects reconstruction, registration, and feature matching in minimally invasive surgery. We present a unified framework that detects and tracks these deformations. A gradient-density formulation locates deformation regions from inflection points of the gradient information-density distribution, without handcrafted priors or manual labels. A self-attention residual encoder-decoder then jointly produces deformation contours and a compact latent representation, and a deformation-aware tracker fuses contour, semantic, and color cues to maintain temporal consistency. Across in-house endoscopic sequences and public segmentation, boundary, and tracking benchmarks, the method attains higher accuracy and more complete, connected contours than gradient operators and existing segmentation models. These stable deformation cues can support feature matching, registration, and overlay stability in augmented and mixed reality surgical navigation.