Enhancing Fast Radio Transient Detection with Mask R-CNN Image Segmentation

TL;DR

The paper tackles the challenge of real-time fast radio transient detection by moving beyond thresholded dedispersed time-series searches to direct analysis in the DM-time plane. It introduces a Mask R-CNN–based pipeline that jointly classifies, localizes, and segments bow-tie-shaped S/N degradation patterns, trained on simulated bursts injected into MeerKAT noise. A novel dedispersion strategy and carefully designed, burst-width–dependent search windows enable near real-time processing and substantially reduce candidate overload while preserving sensitivity. Real-data tests on pulsars and FRBs demonstrate high detection rates, including some sub-threshold events, confirming the approach’s robustness and practical impact for current and future surveys.

Abstract

Traditionally, fast radio transient searches are conducted on dedispersed time series using thresholding techniques based on the statistical properties of the data. However, peaks in dedispersed time series do not directly provide information on the nature of the source. In the DM-time domain, the S/N variation of real, dispersed astrophysical signals forms a characteristic bow tie shape, whereas radio frequency interference (RFI) can take multiple different forms. We have developed a method that bypasses the thresholding step of traditional single-pulse searches in favour of a direct DM-time domain image analysis. The backbone of our pipeline is a Mask R-CNN, a deep learning model designed for object detection, enabling it to identify the bow tie signature and distinguish real sources from RFI. Previous deep learning models often include a snippet of the DM-time domain in their input. We have trained the model on simulated bursts injected on top of real MeerKAT noise observations. We tested the model on MeerKAT follow-up observations of the repeater FRB20240114A and we were able to recover all bursts with a signal-to-noise above the traditional threshold, while detecting two bursts that were fainter. Our new approach considerably reduces the number of candidates above a nominal threshold while being capable of running in real time for typical surveys. We also propose a modified version of the traditional dedispersion plan optimised for this method.

Figures (11)

• Figure 1: The top row displays frequency-time plots for different sources, the middle row shows the DM-time transform of these sources, and the bottom row presents the frequency-summed time series at the point of maximum S/N. The first column depicts a broadband burst dispersed according to Equation \ref{['eq:time-dispersion']}, simulating an ideal transient source. The second column illustrates a linear frequency-drifting source, representing a type of RFI. The third column shows narrowband RFI at two random frequency channels, a common occurrence in radio transient searches. The fourth column represents a narrowband, intermitent burst following the dispersive relation expected from astrophysical sources, illustrating a more complex emission scenario than the ideal case in the first column.
• Figure 2: DM-time transform of a simulated burst and contour lines at different S/N drop points. The distance between the furthest point and the central DM-time point can be used as a reference to indicate the minimum amount of signal that needs to be present in the image to be detected by our pipeline, as shown by the dashed line boxes.
• Figure 3: Smearing time as a function of DM for a dedispersion scheme obtained from our pipeline (top) and the DDplan script from PRESTO (bottom). The green, orange, and blue lines show the intrachannel, DM step, and sampling time smearing respectively. The solid purple line indicates the sum of all the smearing effects. The dotted line shows the width values used for the search.
• Figure 4: A workflow diagram of our pipeline, showing the classification, localisation and segmentation mask prediction steps. The DM-time transform acts as an input to the algorithm. Feature maps are obtained using a backbone CNN, which are up-sampled by the FPN to obtain high semantics across different scales. The RPN proposes regions of interest across the image where relevant objects may be present, performing a light-weight classification. Regions with the highest score are selected and aligned with the original feature map by the RoI Align network. From these RoI, a segmentation mask is obtained through a CNN with an up-sampling layer at the end, and a classification score and bounding box coordinates are obtained through a series of fully connected layers, producing the output of the algorithm.
• Figure 5: Inference time of the Mask R-CNN model for one second of filterbank data across all DM trials (dots) and memory allocated in the GPU (triangles) shown for different input image sizes. The dashed line shows the boundary for real-time processing.