Advancements in artificial intelligence (AI) have revolutionized the field of microscopic data analysis. However, as AI models become more complex, they require substantial computing power and energy consumption. To address this issue, researchers from Leibniz-Institut für Analytische Wissenschaften (ISAS) and Peking University have developed an open-source compression software called EfficientBioAI. This user-friendly toolbox allows scientists to run existing bioimaging AI models more efficiently and with significantly lower energy consumption.
Modern microscopy techniques generate large volumes of high-resolution images, with individual data sets often containing thousands of images. To reliably analyze these data sets, scientists rely on AI-supported software. Unfortunately, as AI models become more complex, the processing time for images increases. This high latency leads to increased computing power and higher energy consumption. Dr. Jianxu Chen, the head of the AMBIOM—Analysis of Microscopic BIOMedical Images research group at ISAS, highlights the importance of reducing latency in image analysis, especially on devices with limited computing power.
EfficientBioAI employs model compression techniques widely used in digital image processing and AI to make AI models lighter and more energy-efficient. By reducing memory consumption and speeding up model inference, the software significantly decreases energy consumption. Techniques such as pruning, which removes excess nodes from the neural network, are used to achieve these results. The researchers aimed to develop a simple and ready-to-use solution for applying these techniques to common AI tools in bioimaging.
The researchers tested EfficientBioAI on several real-life applications, with different hardware setups and various bioimaging analysis tasks. The results showed that the compression techniques significantly reduced latency and cut energy consumption by 12.5% to 80.6%. Importantly, the compression did not compromise the accuracy of the models. Chen provides an illustrative example using the widely used CellPose model. If a thousand users were to compress the model using EfficientBioAI and apply it to the Jump Target ORF dataset (which consists of around one million microscope images of cells), the energy savings would be equivalent to the emissions from a car journey of approximately 7,300 miles (11,750 kilometers).
EfficientBioAI aims to make neural networks in bioimaging more efficient without sacrificing accuracy. The software is user-friendly and can be seamlessly integrated into existing PyTorch libraries, an open-source program library for the Python programming language. For some commonly used models like Cellpose, researchers can use the software without having to modify the code. To accommodate specific change requests, the research group provides demos and tutorials. With just a few lines of modified code, EfficientBioAI can be applied to customized AI models.
EfficientBioAI is an open-source compression software designed specifically for AI models in bioimaging. While it is currently available for Linux (Ubuntu 20.04, Debian 10) and Windows 10, the researchers are working on making it accessible for MacOS as well. The software focuses on improving the inference efficiency of pre-trained models rather than increasing the efficiency during the training phase. The researchers continue to develop the toolbox and expand its capabilities to benefit scientists in biomedical research.
*Note:
1. Source: Coherent Market Insights, Public sources, Desk research
2. We have leveraged AI tools to mine information and compile it
