Abstract Bacterial cellulose (BC) holds broad application prospects in fields such as tissue engineering, flexible electronics, medical dressings, and food packaging due to its high crystallinity, excellent mechanical properties, and biocompatibility. However, significant variations exist among bacterial strains in terms of yield and structural properties. Existing screening methods rely on endpoint yield measurements and empirical judgment, resulting in lengthy screening cycles (≥7–14 days), low efficiency, and difficulty in identifying high-potential strains at an early stage. To address this bottleneck, this study proposes an intelligent screening model for bacterial cellulose-producing strains based on hyperspectral microscopic imaging (HMI) and a multi-modal deep fusion strategy, enabling efficient screening from the colony level to the single-cell scale. This model integrates spectral, image, and cell morphology features, employing a multi-branch architecture combining LSTM, 1D-CNN, and 2D-CNN for deep feature extraction. High-precision prediction is achieved through feature concatenation and Softmax classification. Experimental results demonstrate that the model achieves 98.4 % classification accuracy on the test set, significantly outperforming traditional PCA-SVM and single-modal CNN approaches. It enables rapid identification of high-yield strains during early fermentation stages, substantially shortening screening cycles and reducing experimental costs.
Yan et al. (Thu,) studied this question.