Deep Learning for Isotope Identification and Quantification using Atom Trap Trace Analysis

Atom Trap Trace Analysis (ATTA) has proven to be a valuable technique for detecting rare gas radioisotopes. By tuning the frequency of a laser to the resonance of a desired isotope, only atoms of that isotope are captured by a magneto-optical trap (MOT) and detected by observing its fluorescence with a CCD camera. Existing approaches rely on manually selecting a region-of-interest (ROI) of image pixels where the atoms’ fluorescence signal occurs and then integrating the count in the corresponding region to determine the atom count. Querying for a specific isotope typically produces thousands of images so this manual analysis can be time-consuming and laborious. To automate this process, image analysis algorithms need to handle challenges like scattered light or detector noise, fleeting atoms due to CCD exposure time, and spurious camera events due to cosmic rays and x-rays. Sky Park Labs proposes DeepATTA (Deep Learning for isotope identification and quantification using Atom Trap Trace Analysis), a software suite comprised of machine learning algorithms which will automatically identify and quantify atoms in images produced by ATTA systems in near real-time. DeepATTA employs feature learning techniques to learn a latent representation from unlabeled ATTA images. This representation serves as the backbone for the image analysis pipeline. The analysis pipeline includes atom detection, image denoising, and event classification, to provide efficient and accurate atom counting. The atom detection module uses a machine learning-based detection framework built upon the learned feature to automatically detect and segment the image pixels corresponding to the atoms’ fluorescence signal. These regions are then processed by an image denoising module to remove residual noise that occur in extreme low abundance detection. The resulting image can then be integrated to calculate an accurate atom count. Specific events require specialized processing to yield an accurate atom count. To that end, the event classification module flags the occurrence of specific events, such as high atomic concentrations and spurious events, for additional management. Since spurious events constitute a rare event, DeepATTA learns a model of “normal” behavior exclusively on fault-free, normal data and flags any deviations as potential spurious events. The proposed system represents a substantial advancement in the state-of-the-art and will significantly enhance the turnaround time of sample analysis, moving closer to near real-time monitoring of rare gas radionuclides.