AI-driven analysis of SEM images of thin layers of volcanic rocks: a test with Segment Anything for Microscopy

In recent decades, image analysis has been consolidated as a standard quantitative method. However, with a few exceptions, this analysis has predominantly relied on laborious and highly dependent human expert intervention. Consequently, physicochemical methods, which can be more readily automated for the analysis of large sample sets, have been the most commonly employed approach to provide scientific evidence. This tendency is particularly pronounced in the field of geosciences, where the analysis of light and SEM microscopy images of thin rock layers is informative but can only be performed on a limited number of samples, thereby compromising conclusions at larger scales. In recent years, advancements in artificial intelligence have elevated image analysis to a new level by automating human interpretation, thereby enabling the processing of a greater number of samples. In this study, we examine the Segment Anything for Microscopy (micro-sam) package, which is based on the widely used deep learning tool Segment Anything Model (SAM), to assess its practical application in the analysis of SEM images of thin layers of volcanic rock samples. 

To this end, we have first conducted a grid search of the best SAM parameter values using a set of three SEM images, exploring the impact of the different SAM parameter sets on automatic mask generation. We generated a data set comprising more than 300 objects per image by interactive delineation and identification ("labeling"), and used this data set to evaluate the results and identify the best sets of parameter values for each image, as well as common sets that provided good results across all three images. A common set of parameter values was then used to compare the results obtained from the three available SAM models. The findings of this study indicate that two distinct sets of parameter values are particularly noteworthy. The first set leads to maximized object detection, which is intended to be subsequently used for automatic instance segmentation through deep-learning methods. The second set produces severe object over-segmentation with a very low error rate, making it useful for subsequent classification. Furthermore, we have investigated the micro-sam capabilities of custom fine-tuning by employing our labeled objects as a training set. 

The preliminary findings indicate that deep-learning methodologies, such as micro-sam, can be efficiently implemented for the analysis of SEM images of thin layers of volcanic rock samples. This approach will lead to a substantial increase in the number of analyzed images, provided that appropriately labeled objects are fed to the system. This strategy notably enhances the cost-efficiency of the time invested by experts. In alignment with current practices in related domains, experts in the analysis of these images should collaborate in a concerted manner to generate shared training sets and artificial intelligence models.

This research was partially supported by the HYDROCAL (PID2020-114876GB-I00) grant funded by MICIU/AEI/10.13039/501100011033.