During prostate cancer brachytherapy, catheters are inserted into a patient's prostate for a highly localized radiation treatment. Accurately placed catheters are critical for successful treatment and ultrasound images are taken throughout the procedure to verify their exact positions. However, manually locating catheters on ultrasound images is extremely di cult, time consuming, and happens while the catheters are still in the patient. A fully automatic solution could signi cantly reduce procedure time and potentially even improve the precision. This thesis introduces a novel approach that segments 2D ultrasound images using the successful U-Net architecture to determine catheter candidates. These candidates are then extracted and Fourier Transformed into the frequency domain. De-convolution is performed directly in the frequency domain to reconstruct a number of frequency coe cients and remove noise. Additional features are calculated from the frequency coe cients to supplement the determined U-Net con dence and candidate location. Altogether, the features from each catheter candidate are classi ed by AdaBoost.