Deep Learning-Based Supervised and Semisupervised Segmentation of Confocal and SEM Micrographs for Microstructural Characterization of Plant Protein Gels

Microscopy-based analysis of plant protein gels offers essential insights into their structural organization and functional performance. However, quantitative image segmentation is often constrained by the need for extensive manual annotations. In this study, we developed a semisupervised learning (SSL) framework to segment confocal and scanning electron microscopy (SEM) images of plant-based protein gels using only 10% labeled data combined with pseudo-labeling. The SSL model achieved segmentation accuracy [assessed using Intersection over Union (IoU) and Dice coefficients] comparable to fully supervised learning (SL) models trained on complete annotations, particularly for SEM images (SSL: IoU = 0.83, Dice = 0.91; SL: IoU = 0.84, Dice = 0.91). For confocal images, SSL reached an IoU of 0.73 and Dice of 0.85, versus 0.82 and 0.90 from SL. Structural metrics such as fractal dimension and protein aggregation area were extracted from both ground truth and predicted masks, showing strong correlations and demonstrating the robustness of the SSL model. These descriptors are inherently scale-invariant, enabling reliable comparisons across imaging conditions without magnification calibration. This work presents a high-throughput, annotation-efficient solution for food microstructure analysis. The SSL framework, coupled with scale-invariant structural descriptors, holds promise for broader applications in material characterization and offers a foundation for incorporating size-dependent features in future work.