DLSIA: Deep Learning for Scientific Image Analysis

Area of Science:

• Scientific image analysis
• Machine learning
• Data processing

Background:

• Increasing scale and complexity of experimental data necessitate advanced image analysis tools.
• Traditional methods struggle with large, complex scientific image datasets.
• Need for accessible and customizable machine learning solutions in research.

Purpose of the Study:

• Introduce Deep Learning for Scientific Image Analysis (DLSIA), a Python library for scientific image analysis.
• Provide scientists with customizable Convolutional Neural Network (CNN) architectures for diverse image analysis tasks.
• Simplify the application of deep learning in scientific research.

Main Methods:

• Developed DLSIA, a Python library featuring customizable CNN architectures like autoencoders, U-Nets, and Mixed-Scale Dense Networks (MSDNets).
• Introduced Sparse Mixed-Scale Networks (SMSNets) using random graphs, sparse connections, and dilated convolutions.
• Employed DLSIA networks and training scripts for applications including inpainting, 3D fiber segmentation, and data compression/clustering.

Main Results:

• Demonstrated DLSIA's utility in inpainting X-ray scattering data using U-Nets and MSDNets.
• Successfully segmented 3D fibers in concrete using an ensemble of SMSNets.
• Showcased autoencoder latent spaces for effective data compression and clustering.

Conclusions:

• DLSIA provides accessible CNN construction, abstracting complexities for scientists.
• The library enables tailored machine learning approaches, accelerating scientific discovery.
• Facilitates interdisciplinary collaboration and advances scientific image analysis research.