In brachytherapy of prostate cancer treatment, needles, called catheters, are inserted into the prostate of a patient. The accuracy of the catheters’ position is critical for the effectiveness of the treatment. Ultrasound images are commonly used to measure their positions. However, scattering of ultrasound signal in human tissue produces echoes of catheters that are sometimes, to the human eye, difficult to distinguish. The primary goal of this research is to investigate a novel approach for distinguishing ultrasound image regions corresponding to actual objects from the regions corresponding to their echoes. One of the challenges in ultrasound image analysis is the level of noise that is significantly higher than for other types of medical images, such as CT and MRI. Analysis performed directly on the intensity content of the regions of interest is called the spatial domain approach. Due to the complexity of the problem, Machine Learning has recently become an attractive technique to tackle this problem. On the other hand, it is worth to investigate frequency domain approaches due to their notable success in signal processing, e.g., in voice recognition. To compute the Fourier transform of a local image region, a window function is applied to mask off the remaining area of the image. The classic method for local frequency analysis is the Gabor transform that employs a Gaussian function as the window mask. However, the frequency coefficients corresponding to the intensity content in the region of interest are tangled together (through convolution) with the frequency coefficients of the Gaussian window in the frequency domain. A de-convolution algorithm, called Probing Detector, is utilized to reconstruct the frequency coefficients corresponding to the image content iteratively in the order of magnitude of the coefficients. This research proposes a set of novel feature vectors based on the reconstructed frequency coefficients in the frequency domain. These features are used as the input data to a Neural Network classifier named Probing FCNN. The features derived from the frequency domain can potentially be more robust to noise in the image. At the same time, they can reduce significantly the sample size and training time required by the learning process. The results from our initial experiments are very promising.