Biomedical
Accelerating segmentation of fossil CT scans through Deep Learning
Peer Reviewed
Doi: http://dx.doi.org/10.1038/s41598-024-71245-1
Abstract
Recent developments in Deep Learning have opened the possibility for automated segmentation of large and highly detailed CT scan datasets of fossil material. However, previous methodologies have required large amounts of training data to reliably extract complex skeletal structures. Here we present a method for automated Deep Learning segmentation to obtain high-fidelity 3D models of fossils digitally extracted from the surrounding rock, training the model with less than 1%-2% of the total CT dataset. This workflow has the capacity to revolutionise the use of Deep Learning to significantly reduce the processing time of such data and boost the availability of segmented CT-scanned fossil material for future research outputs. Our final Unet segmentation model achieved a validation Dice similarity of 0.96.
Key Questions
What is the main focus of the study?
The study focuses on developing a Deep Learning method to automate the segmentation of fossil CT scan data, aiming to reduce processing time and improve the availability of segmented fossil material for research.
How does the proposed Deep Learning method improve upon previous methodologies?
The proposed method requires less than 1%-2% of the total CT dataset for training, significantly reducing the amount of manual input needed compared to previous methods that required larger training datasets.
What were the results of the Deep Learning model's performance?
The final U-Net segmentation model achieved a validation Dice similarity coefficient of 0.96, indicating high accuracy in segmenting fossil material from the surrounding matrix.
What is the significance of this research for future studies?
This research has the potential to revolutionize the processing of CT scan data in paleontology by significantly reducing segmentation time, thereby accelerating the study of fossil specimens and facilitating more efficient data sharing among researchers.
Doi: http://dx.doi.org/10.1038/s41598-024-71245-1