Abstract Radiofrequency ablation is a minimally invasive technique for treating liver or kidney tumors, where precisely inserted ablation probes destroy the tumor tissue with induced heat around their tips. Automatic segmentation of needles from the abdominal computed tomography (CT) scans enables large-scale procedure evaluation and creates a baseline ground truth for automated planning approaches. Our method for needle segmentation enables the detection of multiple needles from unlabeled CT scans by sequential combination of two approaches, a classical method and a neural network. The former method detects intensity-prominent needles in 3D. The latter focuses on the detection of needles with lower intensity, that remained undetected by the classical method. A well-known encoder-decoder approach on axial slices of the CT scan gives a segmentation mask for the needles prominent in slices. To train the network, we create a synthetic dataset, adding needles to clean abdominal CT slices. Our network trained on a small synthetic dataset is able to generalize to clinical CT scans. The performance of the combined method is evaluated on ten anonymous patients from the Medical University of Innsbruck each with up to 18 needles inserted in the liver without any template.