Automating the Segmentation of X-ray Images with Deep Neural Networks
In an era of rapid advancements in X-ray physics and the growing capabilities of X-ray synchrotron sources, the analysis of tomographic X-ray datasets has become increasingly critical in various scientific, medical, and industrial applications. However, once the raw data are collected, the manual segmentation of these images is a time-consuming and error-prone process, often plagued by uncertainties and subjectivity. Automating the segmentation process becomes imperative to keep pace with data acquisition rates and to ensure timely scientific discoveries and industrial insights. To address this challenge, in this project, we plan to leverage the power of deep neural networks to automate the segmentation of ptychographic X-ray images, removing the need for human intervention and significantly expediting the analysis process.
The primary objective is the development and training of a deep neural network, based on existing architectures, commonly used for other computer vision tasks (e.g. UNet, VGGnet, etc.). The training dataset consists of real-world X-ray images (raw and segmented), and the results will be benchmarked against manually labeled datasets. The project will be supervised by Salvatore De Angelis ([email protected]) and Peter Stanley Jørgensen ([email protected]).
More info can be found in the Project’s GitHub repository:https://github.com/sdea/AIStudentProjects/edit/main/README.md
10 Terabytes pr day
Syncrotons
Too much data for manuel segmentation
Use machine learning to segment many images
Real data from research
Three phases
Best quality datasets
16 bit greyscale
They are labeled 8 bit values 1, 2, 3
Threshholding, problem is 1000 of these values, they are absolutes