Augmented Synthetic Dataset with Structured Light to Develop Ai-Based Methods for Breast Depth Estimation

Breast interventions are common healthcare procedures that normally require experienced professionals, expensive setups, and high execution times. With the evolution of robot-assisted technologies and image analysis algorithms, new methodologies can be implemented to facilitate the interventions in this area. To enable the introduction of robot-assisted approaches for breast procedures, strategies with real-time capacity and high precision for 3D breast shape estimation are required. In this paper, it is proposed to fuse the structured light (SL) and deep learning (DL) techniques to perform the depth estimation of the breast shape with high precision. First, multiple synthetic datasets of breasts with different printed patterns, resembling the SL technique, are created. Thus, it is possible to take advantage of the pattern's deformation induced by the breast surface in order to improve the quality of the depth information and to study the most suitable design. Then, distinct DL architectures, taken from the literature, were implemented to estimate the breast shape from the created datasets and study the DL architectures’ influence on depth estimation. The results obtained with the introduction of a yellow grid pattern, composed of thin stripes, fused with the DenseNet-161 architecture achieved the best results. Overall, the current study demonstrated the potential of the proposed practice for breast depth estimation or other human body parts in the future when we rely exclusively on 2D images.