Abstract-The localization and identification of vertebrae in spinal CT images plays an important role in many clinical applications, such as spinal disease diagnosis, surgery planning, and post-surgery assessment. However, automatic vertebrae localization presents numerous challenges due to partial visibility, appearance similarity of different vertebrae, varying data quality, and the presence of pathologies. Most existing methods require prior information on which vertebrae are present in a scan, and perform poorly on pathological cases, making them of little practical value. In this paper we describe three novel types of local information descriptors which are used to build more complex contextual features, and train a random forest classifier. The three features are progressively more complex, systematically addressing a greater number of limitations of the current state of the art.

Context. Extragalactic radio continuum surveys play an increasingly more important role in galaxy evolution and cosmology studies. While radio galaxies and radio quasars dominate at the bright end, star-forming galaxies (SFGs) and radio-quiet active galactic nuclei (AGNs) are more common at fainter flux densities. Aims. Our aim is to develop a machine-learning classifier that can efficiently and reliably separate AGNs and SFGs in radio continuum surveys. Methods. We performed a supervised classification of SFGs versus AGNs using the light gradient boosting machine (LGBM) on three LOFAR Deep Fields (Lockman Hole, Boötes, and ELAIS-N1), which benefit from a wide range of high-quality multi-wavelength data and classification labels derived from extensive spectral energy distribution (SED) analyses. Results. Our trained model has a precision of 0.92±0.01 and a recall of 0.87±0.02 for SFGs. For AGNs, the model performs slightly worse, with a precision of 0.87±0.02 and a recall of 0.78±0.02. These results demonstrate that our trained model can successfully reproduce the classification labels derived from a detailed SED analysis. The model performance decreases towards higher redshifts, which is mainly due to smaller training sample sizes. To make the classifier more adaptable to other radio galaxy surveys, we also investigate how our classifier performs with a poorer multi-wavelength sampling of the SED. In particular, we find that the far-infrared and radio bands are of great importance. We also find that a higher signal-to-noise ratio in some photometric bands leads to a significant boost in the model performance. In addition to using the 150 MHz radio data, our model can also be used with 1.4 GHz radio data. Converting 1.4 GHz to 150 MHz radio data reduces the performance by ~4% in precision and ~3% in recall.

No abstract