Weakly-Supervised Hair SEM Microscope Image Segmentation Using a Priori Structure Information

• DOI: 10.5566/ias.3406
• Publication Date: 2024-11-22T23:00:19Z
• Authors: Xiaohu Liu; Thérèse Baldeweck; Pierre Dupuis; Thomas Bornschloegl; Etienne Decencière; Beatriz Marcotegui
• Keywords: Sobel operator; Interpretability

The analysis of microscopic images hair has a wide range applications in the domains cosmetics, healthcare and forensics. segmentation represents initial step automatic large-scale quantitative hair. It ensures that subsequent measurements are performed an appropriate area corresponding to hair, avoiding artifacts near boundaries. This process can be time-consuming, tedious susceptible subjective errors when conducted by human operator. Deep learning methods represent promising solution; however, obtaining pixel-level accurate masks is costly process. paper presents novel weakly supervised pipeline for SEM (Scanning Electron Microscope) images, which requires only simple image-level annotations training. proposed method incorporates Radon transform, Sobel operator Boundary Discrimination Module (BD-module) estimation presence was evaluated on recently collected dataset (429 images). Furthermore, it benchmarked with including Unet Segment Anything Model (SAM). results demonstrated mean Hausdorff Distance improvement over 30% standard deviation 50% comparison SAM. Moreover, we additional refinement modules address boundary nonlinear cases Grad-CAM enhance interpretability BD-module. Additionally, quality metric based gradient map self-quality assessment. accessible research community open-source format.