9. Annual MT Meeting

Robust Image Descriptor for Machine Learning based Data Reduction in Serial Crystallography

Not scheduled

2h 30m

Gaede Lecture Hall (Bldg 30.22) (Karlsruhe Institute of Technology (KIT))

Poster without speed talk Data Management and Analysis Poster session

Speaker

Dr Vahid Rahmani (FS-DS (Detektorsysteme))

Description

Serial crystallography experiments at synchrotron and X-ray Free Electron Laser (XFEL) sources produce large datasets, consisting of multi-megapixel images acquired at a high frame rate with detectors such as AGIPD. However, only a small percentage of the images are useful for downstream analysis, because most X-ray pulses miss the target (protein nanocrystals in a liquid jet). Our goal is to develop computationally-efficient machine learning methods for rejecting bad images on-the-fly. We have developed a pipeline for this, consisting of a real-time feature extraction algorithm called Modified and Parallelized FAST (MP-FAST), an image descriptor, and a machine learning classifier. For parallelizing the primary operations of the proposed pipeline, we have implemented and compared the performance of CPU, GPU, and FPGA processors. Finally, we evaluated our MP-FAST-based image classification using a Multi-Layer Perceptron (MLP) on various datasets, including both synthetic and experimental data.