An integrated deep learning framework for robust multi-view wound modeling in in-situ bioprinting

This is the authors' abstract. We don't add key points for this paper.

In-situ three-dimensional (3D) skin bioprinting promises effective repair of irregular defects, but clinical translation faces two major bottlenecks: robust 3D wound reconstruction under dynamic conditions, and the scarcity of clinical data for training generalizable models. This paper presents a deep learning-driven framework integrating multi-view reconstruction and hierarchical model generation to address these challenges. We propose three key innovations: (1) a 3D Coarse-to-Refined Reconstruction Network with Transformer (3D-C2RT) that integrates confidence-guided multi-view attention to decouple dynamic clinical interference from reliable geometric features; (2) progressive transfer learning adapting models from synthetic to target domains, effectively mitigating data scarcity; and (3) a geometric prior-guided point cloud completion network generating consistent multi-layered models. The framework automatically produces printable layered models from multi-view red-green-blue (RGB) images, providing precise inputs for in-situ bioprinting. Experiments on synthetic 3D datasets demonstrate its superiority, where 3D-C2RT achieves an average Chamfer Distance (CD) of 0.392, showing highly competitive performance against state-of-the-art methods, with the strongest robustness under degraded inputs. Under simulated dynamic interference, the reconstruction error remains stable at 0.67 mm (mm), outperforming a commercial laser scanner by 78.7%. On a 204-case clinical dataset, transfer learning reduces CD by 15.3% and cuts reconstruction time by 90.1% versus traditional structure from motion (SfM) methods. In-vivo rat experiments indicate that the generated multi-layered models, when used for in-situ printing, achieve a 100% wound closure rate by Day 21, surpassing conventional treatments. This work demonstrates a promising pathway from wound sensing to printable model generation for automated in-situ skin regeneration.