Segment Every Fossil - A Deep Learning Model Tailored for Automatic Segmentation of Microfossils in Thin Section Images

Microfossil segmentation and identification, an integral step to paleontological and geological research, has historically required laborious manual analysis for segmentation due to their intricate structures and non-trivial segmentation through traditional Machine Learning and Computer Vision techniques. For tackling this, we introduce "Segment Every Fossil" (SEF), a deep learning model specifically designed for the automatic segmentation of a diverse array of microfossils specimens in thin section images. SEF builds upon established foundation models such as the Segment Anything Model (SAM) and Segment Every Grain, enhancing their capabilities through custom U-Net architectures and attention mechanisms. This meticulous development, supported by a vast dataset of over 8,000 labeled images, propels SEF to achieve an impressive 0.91 accuracy and 0.93 loss, surpassing traditional segmentation methods. This is accomplished through a dual-stage segmentation process that utilizes the custom U-Net for initial segmentation, followed by the Segment Anything Model for precise refinement. This ensures accurate delineation of overall fossil shapes and captures the morphology of the specimen at hand. SEF's custom loss function and U-Net architecture heighten its ability to discern subtle fossil boundaries, a crucial factor for precise analysis, while attention mechanisms guide SEF to prioritize and analyze critical regions within microfossils, enabling detailed morphometric measurements of features such as size, area, and grain-matrix ratios. SEF's performance unlocks possibilities for quantitative microfossil analysis, enabling accelerated research workflows and consequently faster and deeper insights into fossil diversity and evolution. With ongoing efforts to reduce dataset size requirements and explore weakly supervised and self-supervised learning techniques, we aim to enhance SEF's capabilities by exploring the reduction of the dataset size required and make automated microfossil analysis accessible to broader research communities with less data available for training Deep Learning models. SEF represents a significant leap forward in microfossil analysis, offering an accurate, efficient, and being a potentially transformative model for micropaleontological discoveries, holding promise for reshaping the landscape of paleontological and geological research, ushering in a new era of automated and accessible microfossil analysis.