Automatic identification of semi-tracks on apatite and mica using a deep learning method

Fission track dating is a widely used thermochronological approach to constrain the thermal history of rocks. Conventionally this approach requires manual identification of fission tracks under the microscope which can be time-consuming and labor-intensive. In this study, we proceed with the newly developed approach to identify fission tracks on transmitted light images based on a deep learning method. In this new approach, we use a convolution neural network (CNN) to extract semi-tracks through image semantic segmentation. Considering the boundary ambiguity inherent in the CNN, we also extract the multi-scale boundary of the images in order to refine the semantic segmentation. We then calculate an area threshold of semi-tracks to determine whether semi tracks are overlapping or not. These non-overlapping tracks are counted directly from the refined semantic segmentation images. For these overlapping tracks, we develop a boundary-superimposed method by using the refined semantic segmentation and the multi-scale boundary images with the help of the reflected-light images to split them before counting. We used 101 images of spontaneous fission tracks and 7 images of induced fission tracks for training with this new approach and tested the resulting convolutional neural networks on 114 spontaneous fission track images and 60 induced fission track images. Most of the test samples show high precision, recall, F1-score, and overall accuracy, highlighting the potential usage of this approach to identify fission tracks automatically.

Automated summary

This study presents a newly developed approach of using deep learning to automatically identify fission tracks on images, eliminating the need for manual identification. The results show high precision and accuracy, indicating the potential usage of this approach in automatically identifying fission tracks.