Detection of HI Holes in NGC 6946 Using Artificial Intelligence: New Candidates from moment-0 Maps

Abstract

The automated detection of HI holes in galaxy disks has been a long-standing challenge in extragalactic radio astronomy and has become increasingly urgent due to the large data volumes generated by current and forthcoming large-area HI surveys. Conventional approaches based on visual inspection are subjective, difficult to reproduce, and fundamentally incompatible with the scale of the data delivered by instruments such as MeerKAT and the Square Kilometer Array. We present a detection pipeline based on a convolutional neural network, which was specifically designed for the identification of HI holes in moment-0 maps and applied to NGC 6946, a nearby spiral galaxy for which one of the best-characterized HI hole catalogs is available. This approach combines Gaussian high-pass and Gabor filtering, as preprocessing steps tailored to the elliptical morphology of HI holes, with a convolutional neural network trained on 48 visually identified holes from the reference catalog of Boomsma et al. The model achieves a precision of 0.997 and a recall of 1.000 on the validation set, and a mean average precision (mAP50) of 83.2% when applied to the full NGC 6946 moment-0 image evaluated against the 69 cataloged holes used for training and validation. The pipeline also recovers 12 candidate HI holes that are not listed in the reference catalog, with eccentricities between 0.88 and 0.94 and aspect ratios between 2.1 and 3.0. The limitations of the proposed approach, such as the small training set and validation on a single-galaxy, are discussed. Our work establishes a reproducible baseline for automated HI hole detection and provides a foundation for future development toward survey-scale application.