Deep Learning-Based Morphological Analysis for Identifying Culture Status and Contractile Motion in Skeletal Muscle Tissues

This study investigates a deep learning-based framework that uses static microscopic images to identify culture status and contractile motion in cultured skeletal muscle tissues. Noninvasive functional evaluation is important for advancements in tissue engineering and biohybrid robotics. However, acquiring large-scale labeled datasets is often difficult owing to the time-consuming nature of functional measurements. To explore the feasibility of image-based evaluation under the data constraints, we implemented a structured image processing and validation pipeline. First, tissue regions were isolated using U 2 -Net, achieving a Dice coefficient of 0.984 to reduce background interference. Second, an ensemble learning strategy was examined via a culture-day identification task using 528 images. A soft-voting ensemble of convolutional neural networks achieved a Macro F1-score of 0.817, indicating that morphological features associated with culture progression could be effectively captured. Additionally, self-supervised pre-training via SimSiam was tested to extract domain-specific representations from unlabeled images. For contractile motion identification, the ImageNet-pre-trained EfficientNetV2-S paired with a random forest classifier attained an accuracy and F1-score of 75.0% and 0.800, respectively. These results suggest that functional properties of cultured tissues can be partially inferred from static morphological characteristics. While the limited size of the dataset and low identification confidence pose challenges, this study provides a baseline methodology for noninvasive, image-based cellular sensing in engineered biological systems.