In the field of biotechnology and medicine, strain screening and bioprocess monitoring are two crucial aspects, but traditional methods are often time-consuming and experience-dependent.

In recent years, the combination of computer vision (CV) and deep learning (DL) technologies has brought automated, high-precision solutions to this field.

Let images assist intelligent decision-making

Computer vision is able to accurately resolve the morphology and characteristics of microorganisms through image segmentation, feature extraction and other techniques. Deep learning models, especially Convolutional Neural Networks (CNNs), on the other hand, learn classification and prediction laws from massive amounts of data, providing powerful analytical capabilities for microbial detection.

For example, the study shows that 31 CNN models (e.g., ResNet, EfficientNet) are used for colony classification, which can be quickly adapted to microbial image tasks through migration learning, greatly reducing the training cost.

Meanwhile, the optimised YOLOv7 model added a small target detection layer and combined with K-means to optimise anchor frame computation, which significantly improved the accuracy of colony counting,mAP@.5 increase 4.5%, suggesting that the adjustment of the model structure plays a key role in performance improvement.

Let computer vision assist strain screening

Traditional strain screening methods rely on microscope observation and manual counting, which are not only inefficient but also subjective and prone to errors.

Firstly, in terms of high-precision strain identification, the CNN-based DIBaS database (containing 660 images of 33 species of bacteria) has a classification accuracy of 97.24%.

In addition, the classification models of colonies such as E. coli and Pseudomonas aeruginosa were successfully constructed from 1252 petri dish images taken by a 12-megapixel mobile phone, which verified the feasibility of using low-cost equipment for efficient strain screening.

Secondly, to address the problem of small sample datasets, the researchers improved the generalisation ability of the model through data augmentation and migration learning.

Let deep learning optimise bioprocess monitoring

In bioprocesses such as fermentation and pharmaceuticals, the growth dynamics of microorganisms directly affect product quality and production efficiency.CV+DL technology can monitor microbial growth in real time and provide data support for the optimisation of production parameters.

On the one hand, through dynamic growth analysis, using a combination of conventional and hyperspectral cameras, the time window for image acquisition after plant transformation can be accurately determined to optimise the productivity of the bioreactor and ensure the efficient operation of the bioprocess.

On the other hand, in terms of contamination source tracking, deep learning analysis of fungal growth images (containing 17,000 indoor/food fungi) can quickly locate the source of contamination and ensure food safety, which is of great significance for the quality control of bioprocesses.

Despite the remarkable results achieved by CV+DL in microbiological testing, there are still some challenges.

Data bottleneck. Most of the studies rely on laboratory home-grown datasets, such as the 38 studies with mostly laboratory data, which limits the generalisation ability of the model.

Model lightweighting. As the number of CNN layers increases, high computational cost becomes a problem, and needs to be deployed in combination with lightweight architectures (e.g., MobileNet) and edge computing to improve the efficiency of the model in practical applications.

Interdisciplinary Collaboration: In-depth collaboration between CV and biology experts can design algorithms that are more relevant to practical needs.