Catheter-based cardiac interventions have substantially improved the efficacy of the treatment. However, catheter navigation still poses several challenges including its poor visualization and localisation. Therefore, various imaging methods have been used to localize the catheter and estimate its shape. The use of ultrasound imaging is superior to other modalities in many aspects, but suffers from poor spatial resolution. Hence, we present a hybrid approach involving deep learning and classical approach to 3D shape estimation of catheters. Our novel approach uses two stages. First, a UNet-3D model is proposed to estimate the catheter centroid in the ultrasound image. Then, an adaptive Kalman Filter (AKF) is used to fuse the points into the 3D world coordinate frame. Simulation studies, phantom and ex-vivo experimentation results demonstrate the robustness of the method to ultrasound noise and extreme configurations (sharp curves).