Utilizing active learning to accelerate segmentation of microstructures with tiny annotation budgets

Non-destructive 3D imaging techniques, such as X-ray nano-holo-tomography, enable the visualization of battery electrodes. Segmenting electrodes into distinct phases is crucial for a comprehensive analysis, yet the process of precise annotation is very labor-intensive. To address this challenge, deep learning methods have been leveraged for automation. However, acquiring a sufficiently large dataset for training deep neural networks that are widely usable remains impractical. A model that is applicable within a dataset but requires very limited human effort to build and use is a viable direction. We propose an active learning framework operating in a semi-supervised setting that minimizes the annotations required by identifying informative training samples at the pixel level. Our approach achieves accuracy comparable to models trained on complete datasets, while utilizing a mere 4% of the data. We demonstrate the effectiveness of our method through a quantitative analysis involving lithium nickel oxide (LNO) electrodes and a user study focusing on graphite electrodes. Our results underscore the potential of active learning to streamline data analysis through efficient model training.