This study investigated the feasibility of deep learning-based segmentation in intra-abdominal kidney ultrasound registration, to enable image-guided robotic-assisted partial nephrectomy (RAPN). Two state-of-the-art models, DeepLabV3+ and SAMUS, were trained and evaluated using a novel intra-abdominal kidney ultrasound (IAKUS) dataset of 2,265 images from 15 RAPN patients. Moreover, a transfer-learning approach was adopted using the publicly available open kidney ultrasound (OKUS) dataset for pre-training. Results showed that SAMUS consistently outperformed DeepLabV3+ across all metrics, achieving an average Dice score of 88.0±2.0% and Hausdorff distance of 13.7 ± 3.8 mm, consistent with literature. SAMUS was pre-trained on ∼ 30k ultrasound images which enabled a zero-shot test, outperforming trained DeepLabV3+ configurations. Furthermore, no measurable difference was seen between OKUS and IAKUS training. Both findings suggest ultrasound specific features may be more important than organ specific features for training, and data diversity may be more important than strict anatomical similarity. The SAMUS model obtained a registration accuracy of 4.3 ± 2.8 mm and inference time of 4.35 fps, in line with literature reported clinical feasibility. The target registration was even improved by an average of 2.6 ± 4.2 mm in 11/13 patients compared with manual-based registration. This proves that deep learning-based registration is not only feasible in a clinical setting but exceeds manual-based registration.