Comparing U-Net and Mask R-CNN Algorithms for Deep Learning-Based Segmentation of Electron Microscopy Images containing cavities for Nuclear Reactor Applications

In recent times, deep learning techniques have been utilized for segmenting transmission electron microscopy (TEM) images. While these tools have been simplified due to open-source software and developments in theory, expert knowledge is still necessary to select optimal model architecture and training setup. Our study compared the effectiveness of two commonly used segmentation frameworks, U-Net and Mask-RCNN, for segmenting TEM images containing helium cavities. Although U-Net produces high-quality segmentation results, it is not as effective at detecting individual objects as Mask R-CNN, which is a more complex algorithm requiring more computational resources. The selection of algorithms should be based on specific use case requirements; Mask R-CNN may be more suitable for detecting individual objects, while U-Net may be better for efficiency concerns.